cudarc use 12.5 instead of 12.8

[a10y]

When running on lambda.ai boxes, we have CUDA Toolkit 12.8 installed. The NSight Compute version that they ship does not seem to inject libs that include several of the symbols cudarc marks as available on 12.8, for example:

(base) ubuntu@132-145-211-165:~/vortex$ RUST_LOG=vortex_cuda=trace,info FLAT_LAYOUT_INLINE_ARRAY_NODE=true ncu ./target/samply/gpu-scan-cli ./vortex-bench/data/tpch/1.0/vortex-file-compressed/lineitem_0.vortex

thread 'main' panicked at /home/ubuntu/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/cudarc-0.18.2/src/driver/sys/mod.rs:24528:18:
Expected symbol in library: DlSym { desc: "/usr/lib/x86_64-linux-gnu/nsight-compute/target/linux-desktop-glibc_2_11_3-x64/./libcuda-injection.so: undefined symbol: cuTensorMapEncodeIm2colWide" }
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
==ERROR== The application returned an error code (101).

cudarc eagerly loads all symbols it believes should be available at startup, even if they aren't used by the program.

We don't use any recently added CUDA symbols, so to keep this compatible with NSight Compute execution, we should downgrade cudarc to build with only the CUDA 12.5 symbols.

This runs without issue.

[codspeed-hq]

Merging this PR will improve performance by 11.36%

⚡ 1 improved benchmark
✅ 953 untouched benchmarks
⏩ 1466 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	map_each[BufferMut<i32>, 128]	858.1 ns	770.6 ns	+11.36%

---

_{Comparing aduffy/cuda-downgrade-feat (51fa515) with develop (c018251)}

Footnotes

1. 1466 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩