[Codegen][DMA] Fix unaligned swizzle offset computation in gather-to-lds lowering (#24241)

The inverse XOR swizzle applied to DMA source offsets was incorrect in
two cases:

1. **Subgroup base offset**: When a subgroup's transfer size is not a
multiple of the swizzle period, different subgroups sharing the same
local offsets but occupying different rows would get identical swizzled
addresses.

Fix: incorporate the subgroup's base offset within the full allocation
before swizzling.

2. **Access-width alignment**: When `elementsPerLane < accessWidth`, the
integer division inside `swizzleOffset` truncates offsets that differ
only within an access-width group to the same value.

Fix: strip the sub-accesswidth remainder before swizzling and restore it
after. This fix is applied directly in `swizzleOffset` for both XOR and
rotate_rows swizzles. While rotate_rows isn't currently used with DMA,
the access-width alignment issue affects both swizzle types.

Both issues caused numerical mismatches for BF16 batch matmuls using DMA
with XOR swizzle enabled.

Assisted-by: Cursor (Claude)

---------

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>