apply 2D blocking to all kernels

[ahgamut]

extracts a bit more speed in the prompt eval time. also fixes some typo errors.

[ahgamut]

somehow GemmStridedBatchedEx produces garbage output when I try to use 2D blocks there.

[ahgamut]

Alright, I'm unable to find the error that stops GemmStridedBatchedEx. It's a bounds check of some kind, I think.

[ahgamut]

I applied the 2D blocking for every kernel used. This gives a boost in both the CLIP and the prompt_eval parts when using llava, but causes a slowdown in the eval part.

The slowdown is because the blocking for the GemmStridedBatchedEx kernel is not optimal, a lot of threads do no work with the current values of BM, BN, BK. GemmStridedBatchedEx does much better when BM = 32, BN = 4, BK = 32 because less threads are wasted.

[ahgamut]

A balancing act will be have BM/BK/BN as template parameters for all the GPU functions, or redefine those macros again before GemmStridedBatchedEx and copy-paste the body of matmul_block2d.

[jart]

Nice! It looks like this gives us a 13% performance boost for GPU inference (both eval and batch eval) for Windows users. Looks good to me.