feat(linked): add FlashInfer sampling provider

[voltjia]

Summary

• Adds a generic TVM-FFI linked-provider transport with installed header, dependent-library, CUDA source, RPATH, and Python distribution version resolution.
• Adds NVIDIA implementation slot 16 for TopKTopPSamplingFromLogits backed by FlashInfer's AOT sampling module.
• Supports fp16, bf16, fp32, and fp64 logits; top_k_first and joint; CPU/CUDA int32/int64 indices; deterministic seed/offset; default or caller workspace; and non-default CUDA streams.
• Requires flashinfer-jit-cache>=0.6.7,<0.7 and apache-tvm-ffi==0.1.10.

Design

The provider composes FlashInfer's TVM-FFI entry points:

• top_k_first: top-k mask -> softmax -> top-p sampling
• joint: softmax -> joint top-k/top-p sampling

Indices are gathered into a contiguous FP32 logits matrix before sampling, preserving repeated/reordered indices and output batches that differ from the unique-logits batch. Two pinned host staging slots and CUDA events preserve asynchronous stream semantics; the provider-owned default workspace is sequenced across streams, while caller-provided workspace follows the existing Handle ownership contract.

The linked DSOs remain external dependencies under the existing linked-provider deployment contract. They are resolved at configure time and referenced through runtime paths; this PR does not bundle FlashInfer or TVM-FFI into InfiniOps wheels.

Validation

• Release/NDEBUG NVIDIA build passed without provider warnings on A100.
• Linked resolver tests: 33 passed.
• FlashInfer slot 16 focused GPU tests: 10 passed.
• Coverage includes both filter orders, host/CUDA indices, int32/int64 output, seed/offset, default workspace, and a non-default-stream nonblocking sentinel.
• Ruff format/check, clang-format 21.1.8 dry-run, and git diff --check passed.
• Hosted build, Ruff, and clang-format checks passed.
• FlashInfer packed ABI was checked from 0.6.7 through 0.6.18rc1; 0.6.6 is incompatible and rejected by the configured PEP 440 constraint.

Known Boundary

An isolated C++ exact-external-workspace probe compiled, but its test container failed at the first standalone cudaMalloc before entering InfiniOps. The Python GPU matrix exercised the provider-owned workspace and stream path; exact external-workspace runtime remains a residual validation gap rather than an observed provider failure.

Dependency

Stacked on #931, which adds exact implementation selection through consumer-owned ops.json. After #931 merges, this PR can be rebased and retargeted to master.