Fix OOB read in Gather kernel via runtime bounds check on coordinate indices

[TristanInSec]

Summary

The Gather kernel uses TFLITE_DCHECK_GE/TFLITE_DCHECK_LT to validate coordinate indices, but these compile to no-ops in release builds (NDEBUG). This allows attacker-controlled indices from an input tensor to cause an out-of-bounds read from heap memory.

Fix

Replace the TFLITE_DCHECK pair at lines 85-86 with a runtime bounds check:

const CoordsT idx = coords_data[batch * coord_size + coord];
if (idx < 0 || idx >= axis_size) {
  return kTfLiteError;
}

This replaces the previous PR #3534 which was closed for not following the Error Handling Guide.

Alignment with Error Handling Guide

Per Section 2 (Phase 2: Execution - Eval):

"Out-of-bounds Memory Access: If an input tensor contains indices or offsets generated at runtime (e.g., in GATHER, STRIDED_SLICE), these should be bounds-checked at runtime using a raw if (!valid) return kTfLiteError;."

Per Section 4 (Preventing Buffer Overflows in Eval):

"If the overflow comes from invalid control data (like an input tensor providing indices), it should be validated in Eval to prevent memory corruption. Use raw returns."

The guide explicitly names GATHER as an example. The sibling gather_nd.cc (line 155) already validates bounds at runtime with a proper error return.

Testing

Verified with ASan-enabled build: crafted model with out-of-range gather index now returns kTfLiteError instead of reading past the input buffer.

Fixes GHSA-cpmh-q338-mv33