[CK Tile] Add transposed tile load implementation, and tests for load_and_convert_tile

[SamiAario-AMD]

Motivation

Mixed precision b/fp16 x fp8 requires a transposed tile load implementation that supports mixed precision using these types. Implement this, use it in load_and_convert_tile, and add a unit test for load_and_convert_tile which covers this functionality.

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[jakpiase]

I have left one minor request, apart from that it does look great.

[aosewski]

Re-reviewed the changes since my 2026-04-27 round, focused on whether the blocking findings landed. They did:

• Layering inversion — fixed. load_tile_transpose.hpp no longer pulls in the ops/elementwise header; the converter is injected as a PassThroughPack_ template arg, with PassThroughPackSelector living in the ops/common layer. Core stays op-free.
• print_matrix.hpp — fixed. Moved under test/, #pragma once, wrapped in namespace ck_tile, real lens[0]/lens[1], ASCII only, and out of the public host.hpp aggregate.
• LDS WAR hazard — fixed. block_sync_lds() now sits between the LDS read and the next iteration's write.
• gfx950 guard — your rebuttal holds: amd_transpose_load_to_vgpr stays behind #if defined(__gfx950__) in buffer_view.hpp, the convert path is CMake-gated to gfx950, and reverting the static_assert makes sense given the new gfx1250 branch.
• Transpose test coverage — on a closer read I was wrong here. The test is an identity round-trip through a transposed LDS view, and the fill is independent-random (not symmetric), so a broken transpose distribution would scramble values and fail the check. It does exercise the layout.

Warnings are largely handled too (vectorized non-transpose path, gfx950 down-cast builtins, integer_divide_ceil tail handling, PassThroughPackSelector generalization, target gating + rationale). What's left is minor/stylistic — dead LoadTranspose arg in the pk branch, inner Problem shadowing, single test shape — fine to leave or fold in later.

LGTM, approving.

Comments addressed.