[Converter:Bugfix] Fix MNNConvert crash on malformed tflite models (#4796)

[fshuang8299]

Summary

MNNConvert -f TFLITE segfaults on eight malformed models reported in #4796 (all reproducers + ASAN traces attached there). Every crash is a value taken straight from the model file and used without a check that stops execution.

Root causes

1. Null builtin_options union: AsConv2DOptions() / AsPool2DOptions() / AsDepthwiseConv2DOptions() / AsTransposeConvOptions() / AsFullyConnectedOptions() return nullptr when the union tag does not match, and the result was dereferenced unconditionally (e.g. fused_activation_function read through null → SEGV at offset 0xc/0x8/0x14).
2. DCHECK is log-and-continue: logkit.h defines DCHECK(x) as CHECK(x) whose body only prints (no abort), so DCHECK(weightShape.size() == 4) fell through into weightShape[0..3] element loads on empty shapes.
3. Unchecked arithmetic: co * kh * kw * ci in int wraps for adversarial dims (product of large-but-positive dims wraps to a small positive value), producing an undersized destination followed by an out-of-bounds write inside convertDataFormatTflite (the WRITE SEGV at TfliteUtils.cpp:107).
4. Unchecked bias buffer: ::memcpy(biasData.data(), biasDataPtr, sizeof(float) * co) reads from the zero page / OOB when the bias buffer is empty or undersized.
5. ConcatSizeComputer: dereferenced inputs[0] on an empty input list (assert commented out) and trusted a rank that can overflow the 9-slot dim array.

Fix

Reject the malformed operator using the converter's idiomatic dstOp->type = MNN::OpType_MAX pattern (liteConverter.cpp:384 turns it into a clean conversion failure with exit code 1), and fail closed everywhere:

• Null-check every builtin_options accessor (Conv2D, DepthwiseConv2D, TransposeConv, Pool2D, FullyConnected).
• Validate weight shape is 4-D, all dims positive, and weightSize computed in int64 does not overflow int — before any allocation or division (also protects the inputShape[3] / ci group calc).
• Bound-check every bias buffer with size() >= sizeof(T) * count before memcpy / vector range-construction (float, INT8 and UINT8 paths).
• convertDataFormatTflite now validates dims and src/dst and returns false instead of looping through null pointers (defense in depth for all 4 call sites).
• ConcatSizeComputer rejects empty inputs, invalid rank (> MNN_MAX_TENSOR_DIM), and out-of-range axis (per input).

Verification

• All 8 reproducers from [bug] MNNConvert dereferences null builtin_options and empty shape vectors on a malformed tflite model #4796: 8/8 SEGV on master → 8/8 clean exit after fix (exit 1 for rejected models).
• AddressSanitizer build: all 8 reproducers + the legal model run with zero ASAN reports.
• No regression: a generated legal tflite model (conv + depthwise + pool + fc + concat, FLOAT32) converts identically to master — op list byte-for-byte identical, only the random model UUID differs.

Fixes #4796

[fshuang8299]

Verification Report

Complete verification of the fix against the 8 reproducers from #4796, plus regression checks. All tests run locally on x86_64 Linux.

1. Crash reproducers (functional)

All 8 files attached to #4796, run with MNNConvert -f TFLITE --modelFile <file> --MNNModel /dev/null --bizCode check:

Master (baseline)	This PR
8/8 SIGSEGV (exit 139)	8/8 clean exit (exit 1 for rejected models)

2. No-regression: legal model conversion

Generated a legal FLOAT32 tflite model (CONV_2D → DEPTHWISE_CONV_2D → AVERAGE_POOL_2D → CONCATENATION, 8 tensors). Converted with master baseline and with this PR:

• Conversion succeeds on both
• Output op list identical (same op types / names / input-output indexes)
• Numeric inference comparison: both converted models run on CPU with the same fixed input → 128 output elements bit-identical (max abs diff = 0)

3. Quantized path (UINT8)

Generated a UINT8-quantized CONV_2D model (with INT32 bias + quantization params). Conversion succeeds identically on baseline and this PR.

4. Memory safety (AddressSanitizer)

ASAN build (-fsanitize=address): all 8 reproducers + the legal model run with zero ASAN reports. The only LeakSanitizer output is a pre-existing 8-byte static registration object (onnxOpConverterRegister<UnaryOnnx>), present on every MNNConvert invocation regardless of this change.

5. MNN test suite

run_test.out (CPU backend, Precision_High): 380/380 passed, 0 failed — covers the touched source/shape/ShapeConcat.cpp (runtime shape computation) and general core regressions.

6. Style checks (matching repo CI)

• Commit message: [Converter:Bugfix] Fix MNNConvert crash on malformed tflite models (#4796) — matches the CI Check Commit Message Format rule
• Code format: checked with git clang-format --diff origin/master using clang-format 17.0.6 (the version the CI Check Changed Lines Format job installs) — 0 issues

Notes on policy choices

• Depthwise INT8 empty/undersized weight → reject the op (OpType_MAX), unlike Conv2D INT8 which skips the weight. The depthwise path assumes a constant weight buffer; without it the op cannot be represented. Rejecting is fail-closed and avoids silently emitting a zero-weight model.
• ConcatSizeComputer rejects rank-0 inputs and out-of-range axis instead of indexing dim[axis] blindly (the old path was itself out of bounds for scalar inputs / bad axis).
• Rejected ops use the converter's existing idiom (dstOp->type = MNN::OpType_MAX), which liteConverter.cpp:384 turns into a clean conversion failure (exit code 1) rather than a crash.