feat(qemu): implement PTO v0.58 tile semantics and TSORT

[jiale-wangOwO]

Summary

Implement PTO v0.58 tile semantics in Linx QEMU for the shared-tile,
CUBE/TEPL/TLSU, and low-precision numeric paths. This PR consolidates the
prior draft work from #60 and #65 (and the original #66 TSORT work) into a
single integrated branch, then rebases that work onto the current
origin/master (PTO ISA 0.58.3).

Rebase onto 0.58.3

• Rebased feat/tsort-v058-contracts-20260821 with
  git rebase -X ours origin/master (commits already upstream dropped).
• Manually reconciled the two changed helper paths in target/linx/helper.c:
  • linx_tile_preflight_tlsu: restore Core-level aggregate capacity
    (quarter, legacy_whole, shared_capacity) for Shared B.IOS.
  • linx_tile_shared_tmov_local_to_shared and
    linx_tile_shared_tmov_shared_to_local: retain the aggregate Core-level
    payload semantics (issue #112/#132) instead of the 0.58.3 per-PE layout.
• hw/linx/virt.c keeps the 0.58.3 ISA identity
  (pto-isa-0.58.3-mode-function-v1).

Shared tile aggregate semantics

• Interpret B.IOS.TSize with the Core-level aggregate table
  512 << (SizeCode-1) for v0.58 SizeCode=1..12; keep Local B.IOT
  capacity PE-local.
• Fixed PE regions; Shared A/B handling for cooperative TMATMUL paths.
• Apply shared aggregate capacity in tile descriptors, TSize derivation, and
  CUBE destination/profiles.

TEPL numeric (TDEQUANT / TQUANT / TREMS / TINSERT / TFMA)

• TDEQUANT: read the integer source dtype from tile_dtype (B.DATR carries
  the FP32 destination type); accept S8/U8/S32 sources via a variable source
  array.
• TQUANT: affine FP32 -> S8/U8 quantization with optional B.IOR scale and
  zero-point (omission selects multiplier 1.0 and zero point 0).
• TREMS / remainder: FP16/BF16 remainder binary ops plus nonzero-divisor
  preflight for these dtypes.
• TINSERT double-source fix; TFMA accepts FP16/BF16/FP32.
• E8M0 and low-precision encode/decode helpers (type 6..14), including
  ReferenceFloatToE8M0 rounding modes.

TLSU (MGATHER / MSCATTER / MGATHER_CAS / GMOV / TLOAD)

• MGATHER.CAS: per-lane atomic compare-and-swap at GM base plus
  signed/unsigned byte displacement; publishes the observed old value;
  supports S32/U32/S64/U64 index types.
• MGATHER(_MASK) / MSCATTER(_MASK) byte-displacement semantics.
• MSCATTER source-only TSize=000 capacity derivation; use the active
  valid extent for bounds checks.
• GMOV and TLOAD Shared PE0 issuer profiles.

Contracts / docs

• Tile-ISA contracts (decode, helper, tile_isa_058, preflight) and source
  binding via linx_tile_operation_source_count_accepted for variable-arity
  operations.
• Refresh the QEMU tile profile matrix.

Validation

• ninja -C build-linx qemu-system-linx64 (PASS)
• python3 tests/linxisa/test_v058_pto_contract.py (23/23 PASS)
• python3 tests/linxisa/test_v0583_contract.py (14/14 PASS)
• build-linx/tests/unit/test-linx-tile-cube-numeric (PASS)
• Full tests/linxisa/test_*.py suite (PASS)
• git diff --check (clean)

Notes

• Supersedes draft PR fix(qemu): use PE-local group matmul dimensions #60 and fix(qemu): honor FPATR output descriptors and raw tile spills #65; the TSORT content originally in feat(qemu): implement PTO v0.58 tile semantics and TSORT #66 is
  preserved on this branch.
• The old fix(qemu): use PE-local group matmul dimensions #60 branch's e590d918 "frame return target" fix is not included;
  it stays on the experimental exp/shared-capacity-lb-semantics-20260819
  branch.
• source_ref remains pinned pending the issue #132 ADR.

CUBE profile matrix refresh

• docs/linxisa/qemu_tile_profile_support.xlsx regenerated from the
  SuperScalarModel status after repartitioning the CUBE profiles by ASL
  numeric class (Float->F32, Signed->S32, Unsigned->U32, plus MX
  no-scale/with-scale). The previous AccType meta-rule is no longer a
  workbook column.

CUBE matrix refresh (2026-08-24)

• Regenerated docs/linxisa/qemu_tile_profile_support.xlsx after the
  SuperScalarModel status promoted the ordinary TMATMUL Local Float and
  Unsigned class profiles on gfrun.
• QEMU stays UNVERIFIED for those two classes (current QEMU CUBE cannot
  execute any v058 CUBE block); the existing Local Signed matrix pair -> S32
  and all MX class profiles keep their prior both-model PASS state.

[jiale-wangOwO]

Latest 0.58.4 update

Pushed commit 61c6f8213f to this PR branch.

Target spec:

pto-spec@23ca8833fef3f97dbc65beef4924b0b4671cdfdf

This update:

• preserves B.IOS SizeCode bit15 in the internal packed Shared binder;
• adds the unified Shared CUBE compute path for ordinary, BIAS, ACC, MX,
  MX+BIAS, and MX+ACC forms;
• resolves Shared-A/Shared-B and local accumulator/scale/bias source layouts
  according to the latest contract.

Validation:

• ninja -C build-linx qemu-system-linx64 — PASS
• python3 -m unittest tests.linxisa.test_v058_pto_contract — 23/23 PASS

The four-PE carrier reaches Shared TLOAD and CUBE rendezvous but still times
out in its post-CUBE software barrier/finisher stage; that remains a separate
QEMU multi-PE TCG synchronization issue.