fix(attention): Scope FMHA cubin loading by mask type

[suharvest]

What does this PR do?

Scopes context-FMHA cubin loading and caching by the requested attention mask
type.

Previously the load cache key contained only data type and SM. Constructing a
list then attempted to load every matching data-type/SM cubin before the mask
type was considered during function lookup. An unrelated custom-mask cubin
could therefore fail with CUDA_ERROR_INVALID_IMAGE while initializing a
causal model.

The requested mask type now participates in:

• the loader cache key and equality/hash;
• metadata filtering before cubin loading;
• availability and dispatch lookup.

Every caller passes its mask explicitly; the public load function has no
default mask argument. A real custom-mask request still loads the custom-mask
cubin and propagates load failures—there is no fallback to a different mask.

Scope

• Context-FMHA loader/cache implementation and declaration.
• Necessary LLM, ViT, and existing unit-test call sites.
• No cubin, kernel, model, or fallback changes.

Validation

• Source-contract checks verify mask-aware equality, hashing, metadata
  filtering, cache lookup, availability, and dispatch.
• All nine declarations/calls use the explicit three-argument API.
• The loader does not catch or suppress custom-mask load exceptions.
• git diff --check and merge-tree validation pass.

The validation machine has no CUDA/TensorRT toolchain, so native compilation,
the non-custom-mask positive engine case, and a real invalid custom-mask cubin
negative test are still requested from CI/device testing.

Closes #143.