Add optional TBLIS CPU contraction provider

[Ryo-wtnb11]

Summary:

• optional cpu-tblis provider on the CPU backend, with fallback to the existing CPU path when TBLIS is unavailable or not selected
• complex contraction support is wired through the same route
• provider resolution now goes through tenferro-tblis-src, with a pinned TBLIS commit and static linking so the runtime/test/bench binaries do not depend on libtblis.dylib
• docs and the worklog were updated to match the new provider flow

Tests run:

• cargo test -p tenferro-cpu --no-default-features --features cpu-faer,cpu-tblis passed: 220 unit tests, 7 backend capability tests, 3 provider feature tests, 25 runtime error tests, 61 doctests
• cargo check -p tenferro-runtime --no-default-features --features cpu-faer,cpu-tblis passed
• cargo check -p tenferro-linalg --no-default-features --features cpu-faer,cpu-tblis passed
• cargo check -p tenferro-cpu --no-default-features --features cpu-faer,cpu-tblis --benches passed
• cargo check -p tenferro-cpu --no-default-features --features cpu-faer,cpu-tblis,blas-openblas --benches passed after source-provider downloads
• cargo package -p tenferro-tblis-src --allow-dirty --offline passed
• cargo fmt --all --check passed
• git diff --check passed
• otool -L on the test/bench binaries showed no libtblis.dylib dependency

Benchmarks:

• Quick local release benchmarks with threads pinned: TBLIS_NUM_THREADS=1, RAYON_NUM_THREADS=1, OMP_NUM_THREADS=1, and for the BLAS runs OPENBLAS_NUM_THREADS=1
• Small/default BLAS run:
  • f64 matmul 32: faer 3.42 us, BLAS 1.98 us, TBLIS 5.66 us
  • f64 matmul 64: faer 12.71 us, BLAS 19.77 us, TBLIS 19.52 us
  • f64 matmul 128: faer 86.88 us, BLAS 99.27 us, TBLIS 101.31 us
  • c64 lhs-conj matmul 128: faer 385.69 us, BLAS 406.12 us, TBLIS 333.33 us
  • rank-4 interleaved n=8: faer 16.79 us, BLAS 29.64 us, TBLIS 19.08 us
  • rank-5 batched interleaved n=8: faer 62.11 us, BLAS 108.55 us, TBLIS 74.58 us
  • rank-4 row-major view n=8: faer 17.90 us, BLAS 42.66 us, TBLIS 20.13 us
• Large sample:
  • f64 matmul 256: faer 617.38 us, BLAS 748.55 us, TBLIS 645.64 us
  • f64 matmul 512: faer 5.04 ms, BLAS 4.91 ms, TBLIS 4.78 ms
  • c64 lhs-conj matmul 256: faer 2.93 ms, BLAS 2.77 ms, TBLIS 2.39 ms
  • c64 lhs-conj matmul 512: faer 23.37 ms, BLAS 22.08 ms, TBLIS 18.32 ms
  • rank-4 interleaved n=16: faer 672.51 us, BLAS 676.83 us, TBLIS 646.17 us
  • rank-5 batched interleaved n=16: faer 3.03 ms, BLAS 3.10 ms, TBLIS 2.67 ms
  • rank-4 row-major view n=16: faer 935.45 us, BLAS 954.64 us, TBLIS 651.67 us
• Interpretation: TBLIS is not a broad small-size win; it is more favorable on larger complex, interleaved, and view-heavy cases, so this PR does not overstate it as a universal replacement

[Ryo-wtnb11]

Superseded by internal-branch PR #1344 so repository rules review can run on a non-external PR.