fix(ir): state what a symbolic mesh mode and a broadcast operand reach

[zhen8838]

Why

Two places assume a value they are handed is a number, and raise rather than
answer when it is not. Both block the documented workflow of
tilefoundry tutorial optimize on an ordinary program.

• selected_run sorts a mesh level's modes by (stride, extent) and joins
  adjacent ones. A Mesh layout naming an extent a DimVar decides — how a
  chunk-parallel placement over an open sequence length is written — makes the
  strides around it symbolic, and the sort raises
  '<' not supported between instances of 'int' and 'Call' with no source
  location. Its neighbours in the same file already guard on isinstance(..., int). Issue fix(ir): handle a symbolic extent in selected_run #201.
• identity_relations gives every read operand of an elementwise instruction an
  identity of that operand's own rank. A scalar operand therefore yields
  { [] -> [] }, which iterating rejects against the Op's domain, and
  schedule candidates aborts for the whole program: a boundary is asked by 0 coordinates and its Op iterates 2. Adding an epsilon to a reduction is the
  common case, so every normalisation hits it. A lower-rank broadcast operand
  fails the same way. Issue fix(schedule): give an elementwise read the result's rank #200.

What

• ir/types/mesh.py: when any mode of a level carries a symbolic extent or
  stride, selected_run reports the modes the mesh states instead of ordering
  and joining them. A symbolic mode orders against nothing and is adjacent to
  nothing, and two equal runs still compare equal, so covered_by_scope keeps
  recognising a scope as itself.
• visitor_registry/access_relation.py: identity_relations states each read
  operand's broadcast relation with the broadcast_access that HIR Binary
  already uses, so a scalar operand is read once per point the Op walks and a
  shorter one right-aligns. The docstring said a structural operand borrows the
  iteration rank; this is what makes that true for a rank-zero tensor too.
• tests/ir/types/test_mesh.py: a regression test for a run whose modes a
  DimVar decides.

Contract

Neither change alters what a program means or what any already-working program
lowers to. selected_run's result for an all-integer level is unchanged; the
new branch is reachable only where the old code raised. identity_relations
returns the same identity relation whenever every read operand already carries
the result's shape, which is every case that type-checked before. The
instruction's own pattern matching is untouched: an operand it cannot accept is
still refused by shape, dtype or storage, and is now refused with a report
rather than an abort.

Risk

tests/ir, tests/schedule, tests/visitor_registry, tests/analysis,
tests/parser and tests/dsl pass. The full suite's 41 failures are the
pre-existing ones that compile CUDA in this image; the set is identical before
and after the change, verified by running it on both revisions.

The residual risk is in the second change: an instruction whose operands were
previously rejected for an unrelated reason may now be probed further and
reported as refused on a different ground. That is the intended effect — a
candidate report that names the real reason — but it does change the text of
some schedule candidates output.

[zhen8838]

Closing to redo under docs/plans/schedule-broadcast-symbolic-mesh: elementwise relations will share one broadcast-aware identity_relations derived from operand types only, and symbolic mesh scopes will be compared without selected_run. A new PR with a schedule fixture covering #200 and #201 will follow.