metal : remove memory pools

[ggerganov]

ref #15832 (comment)

Instead of allocating buffers from MTLHeap we expand the dst tensors with enough space to write the necessary scratch data.

TODO:

• Better way to keep the get_alloc_size sizes in-sync with the actual implementation (Edit: use simple functions to get the extra sizes)
• Think if this could somehow interfere with the memory ranges logic introduced in metal : allow ops to run concurrently #15929. Might need to start using the extended sizes in the ggml_mem_ranges instead of just using ggml_nbytes

[ggerganov]

@slaren The trick with expanding the allocated size of the tensors seems to work without any problems. Let me know if you have any additional thoughts about utilizing this approach.

[slaren]

LGTM. I suppose there could be an issue where the extra data is not aligned in some cases, but since these operations always return F32, I guess it will always be aligned to at least 4 bytes. Not sure if there could an advantage to using a larger alignment.