Optimize dynamic workflow by pruning node inputs (#6888)

[VishnuVV27]

Tracking issue

Closes #6888

Why are the changes needed?

Dynamic workflows in Flyte can fail or suffer from performance issues when the dynamic closure becomes too large, exceeding gRPC message limits (the "message too large" error). Even if execution succeeds, the Flyte UI becomes sluggish because it has to download and parse these massive blobs. This is primarily caused by the inclusion of all node inputs in the dynamic closure, which is redundant since this data is already persisted in metadata storage.

What changes were proposed in this pull request?

1. Pruning Logic: Added a clearNodeInputs helper in flytepropeller/pkg/controller/nodes/dynamic/dynamic_workflow.go to nullify the Inputs field for nodes in the dynamic closure.
2. Recursive Support: The function specifically handles ArrayNode (Map Tasks) to ensure nested node inputs are also pruned, which are often the largest contributors to payload bloat.
3. Integration: Integrated the pruning step into buildContextualDynamicWorkflow so it triggers just before the closure is cached and sent as an event.

This effectively reduces the "blueprint" size of the dynamic workflow without affecting execution, as nodes continue to read their actual inputs from the defined storage locations.

How was this patch tested?

Added a new test suite TestClearNodeInputs_WithCoreBindings in flytepropeller/pkg/controller/nodes/dynamic/dynamic_workflow_test.go.

• Positive Case: Verified that standard node inputs are set to nil.
• Nested Case: Verified that inputs inside ArrayNode targets are also nullified.
• Verification Command:

  go test -v ./pkg/controller/nodes/dynamic/... -run TestClearNodeInputs_WithCoreBindings