[WIP] Optimizing dynamic alloca ISEL

[easyonaadit]

To get the allocation size for dynamic allocas,
first we scale the max_size by the wave_size,
and then bump SP by this amount.

addr = sp
size = max_size << wave_size
sp = addr + size

Currently, during DAG->DAG I-Sel, the scalar
left-shift and addition operations are being
folded into a v_lshl_add_u32 instruction
and then readfirstlane is used to copy the
value into SP.

This patch enforces the use of scalar operations,
removing the need for a readfirstlane instr.
This makes the code-gen for S-DAG and G-ISEL
uniform.