cpu: aarch64: conv: optimize sve_1x1 kernel

[Ryo-not-rio]

• Removed unnecessary loads
• Updated to use LD_MUL_VL and ST_MUL_VL where possible
• Move add_vl_or_imm to jit_generator.hpp and change add_imm to add_vl_or_imm where appropriate
• Updated to NOT use explicit broadcast in most cases
• Fix potential invalid immediate parameter for ld1rw in brgemm_kernel

Speedups on 16 threads:
c7g: ~1.11
c8g: ~1.11x ~ 1.89x

[Sqvid]

• Apply 1x1 conv optimizations
• Fix potential bug in brgemm_conv kernel

Could you provide a high-level description of what optimisations you made, and what bug you fixed?

[Ryo-not-rio]

All nightly unit tests are passing and a lot of conv regressions are resolved with this patch: https://github.com/uxlfoundation/oneDNN/actions/runs/19906852997

[jondea]

This is really great work, massive speed ups, thank you! I just have some minor comments.

It's probably not worth it at this point, but I'd say that in future, it would have been easier to understand if each optimization had its own commit which explained the reasoning.