Fix uncoalesced global accesses

[danielvegamyhre]

Stacked PRs:

• Multiply with inverse scale in loop inside of dividing to improve perf #12
• ->Fix uncoalesced global accesses #11
• add bfloat16 support for colwise scaling #10
• Fix numerics for colwise scaling #9
• fix scale shapes #8
• update impls of some helpers and add comments #7

---

Fix uncoalesced global accesses

Need scales_colwise to be the shape (num_blocks, columns, 1) so the 'column' dim stride is 1, so we can avoid uncoalesced writes to global memory.

This is because each of the 32 threads in a warp will be computing a scale for a different column of 32 input data values, then each writing that scale to global memory, one per column - so the stride along this col dim should be 1 so writes can be coalesced into a single transaction.

NCU before change

NCU after change