fix(codegen,runtime): honour every topology a Mesh spans

[bigSheep123]

fix(codegen,runtime): honour every topology a Mesh spans

Fixes #66.

What was wrong

Mesh models an ordered topology sequence — topology is the primary level and
topologies the full tuple (shard §5) — but the
emitter wrote only the primary one into the C++ type, and
shard_axis_projection derived the mesh coordinate from
program_id<primary_scope>().

A mesh over cta × thread therefore emitted a type whose topology claims a
16-element domain against a layout of 4096, and every thread of a CTA resolved
to the same mesh coordinate. Each CTA wrote one thread's worth of elements and
left the rest of its rows untouched — 32640 of 32768 elements wrong on the test
added here, with no diagnostic.

The change

• Mesh<TTopo, TMeshLayout> → Mesh<TTopo, TMeshLayout, TMoreTopos...>.
  The primary topology stays first, so every existing single-topology
  instantiation renders and behaves exactly as before. The new linear_id() is
  the mixed-radix position across the pack, coarsest outermost — which is the
  index the mesh layout is already built against
  (cta_id * thread_size + thread_id).
• shard_axis_projection takes mesh_t::linear_id(). get_hier_coord
  already did the hierarchical decomposition; it was simply being handed one
  scope's id.
• The mesh-scope and tensor-view emitters share render_topology and
  extra_topology_args. They had duplicated the rendering, and the alias the
  mesh-scope emitter registers is matched against the tensor-view emitter's
  inline type by string equality, so the two must not drift — that coupling
  is why this is one change rather than two.

Test

tests/integration/test_multi_topology_mesh.py — 16 CTAs × 256 threads, 8
contiguous f32 per thread, gmem → rmem → gmem.

It asserts the full output rather than sampling. That is deliberate: the
untouched elements keep whatever the output buffer already held, so a wrong mesh
coordinate still produces a plausible-looking result and a spot check passes.
The test seeds the output with NaN and the input away from zero so an untouched
element cannot coincidentally look right.

Verified to fail on main before the fix (32640 of 32768 elements wrong) and
pass after.

Verification

pytest tests/ -q     644 passed (main @ e55e4b0)  ->  645 passed (+1 new)

Every .pre-commit-config.yaml hook clean: ruff, spec-rules, spec-refs,
spec-entropy, forward-references, comment-hygiene, no-machine-paths,
english-only, and clang-format over all C++ files.