Add more types to `gpu.shuffle->llvm` lowering

[victor-eds]

Currently, gpu.shuffle->llvm lowering supports just a handful of types: i8, i16, i32, i64, f16, f32, f64. This is missing a couple of types we may wanna handle. Currently, we have to cast original values to supported types in order to use the operation. Extending the operation upstream should help alleviate this need.

Add support for:

• bf16 (by bitcasting to i16 for now)
• i1 (try using SPIR-V builtin or by previous extension to i8)

Note shuffling of other value types like pointers should still be handled in our codebase.

[sommerlukas]

@PietroGhg has started looking into this issue.

[PietroGhg]

Made good progress on this, bf16 is done and i1 coming up well

[PietroGhg]

Should be addressed by llvm/llvm-project#119675

[whitneywhtsang]

Merged in #3236, please verify.

[PietroGhg]

Merged in #3236, please verify.

The new LLVM tag contains the upstream changes for this issue, however it looks like we are still converting e.g. bf16 to f32, so the native bf16 shuffle is not being emitted. I've created #3292 to look into disabling the conversion.

[sommerlukas]

#3292 confirmed that there is currently no need for additional implementation work after the upstream changes, closing this issue.