fix(HwSort): update sorting logic to prioritize valid elements

[ywlcode]

1

HwSort only swapped two entries when both were valid. An invalid entry could block valid entries behind it, so valid entries were not guaranteed to be packed together or sorted.

Example:
in0: valid, ptr = 10
in1: invalid, ptr = X
in2: valid, ptr = 5

The old logic could keep this order:

valid ptr=10, invalid, valid ptr=5

Solution: change the base 2-entry compare-swap to valid-first ordering:

val swap = xVec(1).valid && (!xVec(0).valid || cmp(xVec(1).ptr, xVec(0).ptr))

New behavior: valid entries are moved to the front and sorted by cmp. Equal keys and invalid entries are not guaranteed to be stable.

2

The old 4-way final-stage wiring was incorrect: it compared (0,1) but wrote the result to (1,2), and compared (2,3) but wrote the result to (0,3).

This could corrupt even already sorted inputs:

input = [0, 1, 2, 3]
output = [2, 0, 1, 3]

The fix writes each compare-swap result back to its matching lanes:

tmp3_1 -> row3(0), row3(1)
tmp3_2 -> row3(2), row3(3)