Using `gemv` for `batched_vec`

[AntonOresten]

batched_vec is currently implemented to use batched_mul (which calls batched gemm) only with some extra reshapes.

Some basic benchmarks (on an RTX PRO 6000 Blackwell) suggest batched gemv is sometimes 1-2% faster:

and converges to be similar in the limit:

but in some cases is consistently slightly slower:

Not sure if this has been considered already. The difference isn't huge, but in the cases where there actually is justification to specifically use gemv, it could be nice to have the option.

[CarloLucibello]

how would you implement the interface?

[AntonOresten]

Probably with an alternate batched_vec! dispatch for the CUDA path.

Since benchmarks show batched_gemv is at most 1-2% faster than the current batched_mul approach, and is sometimes slightly slower depending on shape. Closing for now, since it's a memory-bounded op and very similar to GEMM on GPUs anyway.