feat(linked): add FlashInfer sampling provider

[voltjia]

Summary

• Adds NVIDIA implementation slot 16 for TopKTopPSamplingFromLogits, backed by FlashInfer's AOT sampling library through the TVM FFI transport merged in feat(linked): support TVM FFI providers #945.
• Supports fp16, bf16, and fp32 logits; top_k_first and joint; CPU/CUDA int32/int64 indices; deterministic seed/offset; caller-provided workspace; and non-default CUDA streams.
• Keeps linked DSOs external: flashinfer-jit-cache>=0.6.7,<0.7 and apache-tvm-ffi==0.1.10.
• Contains only the provider manifests, implementation, and focused tests; selector and transport changes are already provided by feat(build): select operator implementations from ops.json #931 and feat(linked): support TVM FFI providers #945.

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. Float64 logits are rejected explicitly because this provider cannot preserve their ordering through that FP32 matrix. Float64 top_p remains accepted and is converted toward the interior of (0, 1) so underflow or rounding cannot accidentally disable top-p filtering.

Two pinned host staging slots and CUDA events preserve asynchronous H2D staging. A missing caller workspace is allocated and released per call with cudaMallocAsync / cudaFreeAsync; cached descriptors therefore do not retain shape-sized GPU matrices. CUDA, TVM FFI, metadata, and unsupported-input checks remain active in Release builds.

Validation

Manual NVIDIA validation used the exact #945 tree plus the six blobs from commit 2e69e648; post-#945 master changes are unrelated backend work and are covered by PR CI.

• Resolver/schema tests: 53 passed.
• Release CUDA build: infiniops and Python ops targets passed for SM80.
• FlashInfer implementation 16 GPU tests: 17 passed.
• Dynamic dependencies: readelf found sampling.so and libtvm_ffi.so; GPU-mounted ldd -r had no unresolved dependencies.
• Descriptor cache stress (batch 2..32, vocab 32768): 0 MiB retained device memory after synchronization and allocator cleanup.
• Ruff 0.15.22 check and format check passed; git diff --check passed.

Tested with flashinfer-jit-cache==0.6.7.post3+cu130, apache-tvm-ffi==0.1.10, CUDA 13.1, and an SM80 target.

Dependencies

• feat(build): select operator implementations from ops.json #931 (merged): selects exact implementation slots through consumer-owned ops.json.
• feat(linked): support TVM FFI providers #945 (merged): provides the generic TVM FFI linked-provider transport.