#14826: reimplement l1 data copy

[nathan-TT]

Ticket

#14826

Problem description

Now that the crt reorg has landed (#15094), this reimplements the bespoke memcpy we use to copy from l1 to local memory.

What's changed

1. Reduce insns in the loop. Original loop was 21 insns (3.5 per word), new loop is 10 insns (3.3 per word).

2. Do not use a loop for residue. We only have to handle 0, 1 and 2 cases. A loop is more overhead.

Checklist

• [ YES] Post commit CI passes
• [ YES] Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable)
• Device performance regression CI testing passes (if applicable)
• New/Existing tests provide coverage for changes