Generate simpler LLVM IR for shuffles that recursively become broadcasts

[abadams]

LLVM 18 is causing us to fail the test performance_nested_vectorization_gemm on arm without arm_dot_prod as of a few days ago. The cause seems to be that we generate very complex chains of shuffles for the following piece of IR:

shuffle({some_u8x2}, {0, 1, 0, 1, 0, 1, 0, 1});

The current CodeGen_LLVM Shuffle handler reinterprets the u8x2 as a u16, and creates the following shuffle:

shuffle({reinterpret<u16>(some_u8x2)}, {0, 0, 0, 0});

and then recursively calls the shuffle codegen visitor. The shuffle codegen visitor doesn't have a special case for broadcasts, so it sees this as a degenerate self-interleave, and produces a complex binary tree of shuffles.

This PR instead detects broadcasts of a single lane of a single vector, and uses the existing broadcast handling.

[abadams]

Fixed the thing I was trying to fix. Failures are unrelated.

[abadams]

Remaining failures:

• llvm ICE on fuzz_schedule test on wasm
• llvm ICE on correctness_mul_div and correctness_logical on cuda
• numpy was missing from the venv on mac-worker-3, I believe I have fixed it now