[Bug][Runtime]: context FMHA loads unrelated mask-type cubins and can fail with CUDA_ERROR_INVALID_IMAGE

Describe the bug

The context-FMHA loader in TensorRT-Edge-LLM v0.9.1 / current main
initially filters and caches cubins only by data type and SM. It eagerly calls
cuModuleLoadData for every matching cubin before the requested attention
mask type participates in function selection.

On JetPack 6.2 / CUDA 12.6 / SM87, a non-custom-mask Qwen3.5 engine build
therefore attempts to load two unrelated CUSTOM_MASK cubins that were
generated with CUDA 12.8. Both fail with CUDA_ERROR_INVALID_IMAGE, blocking
an otherwise supported causal/non-custom-mask engine.

The two failing, unused cubins are:

fmha_v2_flash_attention_fp16_fp32_64_128_S_q_k_v_256_custom_mask_sm87.cubin.cpp
fmha_v2_flash_attention_fp16_fp32_64_16_S_q_k_v_256_custom_mask_sm87.cubin.cpp

Nine non-custom-mask SM87 context-FMHA cubins and all required Qwen XQA
cubins load successfully on the same system.

Steps/Code to reproduce bug

Build the unmodified project for SM87 and run llm_build for a Qwen3.5
configuration with vision/custom-mask attention disabled:

./build/examples/llm/llm_build \
  --onnxDir ./qwen35-onnx \
  --engineDir ./qwen35-engine \
  --maxBatchSize 1

The build reaches context-FMHA initialization and fails while loading one of
the unused custom_mask_sm87 modules with
CUDA_ERROR_INVALID_IMAGE.

Current load/cache implementation:
https://github.com/NVIDIA/TensorRT-Edge-LLM/blob/7f061f21f0a581ba234a1e233c9315b89d8e47d6/cpp/kernels/contextAttentionKernels/contextFMHARunner.cpp#L193-L250

Expected behavior

The loader should load and cache only cubins eligible for the requested data
type, SM, and attention mask contract. A causal/non-custom-mask engine should
not load unrelated custom-mask modules. If CUSTOM_MASK is actually
requested and its required cubin is unsupported, the request should continue
to fail loudly rather than silently falling back.

System information (Edge Device)

• Platform: NVIDIA Jetson Orin NX 16 GB
• Software release: JetPack 6.2 / L4T 36.4.3
• CPU architecture: aarch64
• GPU compute capability: SM87
• Total device memory: 16 GB
• Build type: Release
• Library versions:
  • TensorRT Edge-LLM: 7f061f21f0a581ba234a1e233c9315b89d8e47d6
  • CUDA: 12.6
  • TensorRT: 10.3
  • C++ compiler: GCC 11.4
• CMake options:
  • CMAKE_TOOLCHAIN_FILE=cmake/aarch64_linux_toolchain.cmake
  • EMBEDDED_TARGET=jetson-orin
  • TRT_PACKAGE_DIR=/usr

A mask-scoped loader/cache change has completed a fresh non-custom-mask
engine build and direct inference. A follow-up PR will include focused
cache-order coverage and a true custom-mask fail-loud regression.

Start in cpp/kernels/contextAttentionKernels/contextFMHARunner.cpp around lines 193-250, then reproduce the Qwen3.5 non-custom-mask build with the provided llm_build command on SM87. Trace which cubins are loaded before mask selection. Done means non-custom-mask builds avoid unrelated CUSTOM_MASK cubins, while unsupported custom-mask requests still fail loudly; the issue also calls for cache-order and custom-mask regression coverage.