Make API inference_steps defaults model-aware to prevent noisy XL-SFT generations

[Copilot]

API generations were defaulting to inference_steps=8 regardless of model, which is appropriate for turbo but under-sampled for SFT variants and can surface as noisy output. This aligns API behavior with model-specific defaults already used in UI paths.

• Model-aware defaulting in request build path

  • Updated acestep/api/http/release_task_request_builder.py to derive default inference_steps from requested model name when the field is omitted:
    • turbo → 8
    • sft → 50
    • explicit non-turbo/non-sft models → 32
    • no model provided → legacy default 8 (preserved for backward compatibility)
  • Added token-boundary model detection (turbo/sft) to avoid partial-string false positives.

• API docs updated for effective defaults

  • Updated docs/en/API.md and .claude/skills/acestep-docs/api/API.md to document model-aware default behavior for inference_steps.

• Focused regression/unit coverage

  • Expanded acestep/api/http/release_task_request_builder_test.py to cover:
    • SFT/turbo/base/unspecified model defaults
    • token boundary behavior and precedence (turbo over sft when both appear)
    • case-insensitive model handling
    • parser compatibility for seed fallback behavior used by existing parser test fakes

def _default_inference_steps_for_model(model_name: Optional[str]) -> int:
    normalized = (model_name or "").strip().lower()
    if not normalized:
        return 8  # legacy default when model is omitted
    if _has_model_token("turbo", normalized):
        return 8
    if _has_model_token("sft", normalized):
        return 50
    return 32