Fix #443: tensor symbol identity includes dim and rank - #458

Closes #443.

A tensor symbol's identity covered only its name, so the same name at another shape was the same symbol:

Change. symbol_base's operator== / operator< compare the shape after the name, through a requires-guarded detail::symbol_shape helper that yields {dim, rank} for symbols whose base carries them and {0,0} otherwise. Scalar symbols are unaffected.

Extending symbol_base rather than overriding in tensor keeps symbol identity in one place: an override would have to restate the name and hash rules, and the last change to them (#348's name tiebreak) would then have needed mirroring in two files — the maintenance hazard behind this bug family.

The hash is deliberately untouched. It remains a name-only fast reject, so A(3,2) and A(2,2) still hash alike and are separated by the full comparison. Verified unchanged: hash-driven print order (pow(x,2)*y*z*X), tensor_scalar_mul(const,T).hash == T.hash, tensor_pow(T,const).hash == T.hash, and all 27 hash tests.

Tests (SymbolIdentity, beside the #348 cases): TensorShapeDistinguishesSameName covers the four probes plus ordering totality, incomparability of equal symbols, and a 3-entry map; EvaluatorKeepsBindingsPerTensorShape binds both shapes in one tensor_evaluator and asserts each returns its own data, with a trace check through the t2s evaluator.

Negative control: against main's symbol_base.h both tests fail (restored with git show, rebuilt, fresh binary) in Debug and Release; both pass with the fix. Unlike the sibling #348, Release does not crash here — the evaluator dispatches on the bound data's own shape, so the failure is silently wrong numbers.

Suites: Debug 2407/2407, Release 2407/2407 (gcc-14, -Werror minus the pedantic/deprecated-declarations warnings already on main). clang-format-18 clean.

Scope notes. Scalar symbols carry no shape-like state, and there is no t2s symbol type (t2s wraps a scalar expression), so nothing analogous is excluded elsewhere. The parser was never exposed: symbol_table already raises redeclaration_error when a name reappears with a different (rank, dim) — this was reachable only through the C++ API.

Review finding — a second silent wrong result this also fixes. A(3,2) * A(3,4) collapsed to pow(A,2) on main: a rank-2 and a rank-4 tensor merged as one factor, because the like-term path compared symbols by name only. With shape in the identity it correctly stays A*A. Tensor multiplication validates dim but not rank (#438 covers +/-), so that path stays reachable — now pinned by SymbolIdentity.MixedRankFactorsDoNotMerge, which fails against main's identity.

Review fix — shape is compared where the shape lives. The first version put the comparison in symbol_base, which had to detect dim()/rank() with a requires clause: the base of every symbol then knew about state only a tensor carries. tensor now defines its own ==, !=, < and >; equals_same_type/less_than_same_type cast to the derived type before comparing, so these win by exact match and every path reaches them. symbol_base is back to name-only identity with a comment pointing at the extension mechanism.

No virtual was needed: the dispatch already goes through the derived type. Negative control: with tensor.h restored from main, the three shape tests fail while the two cross-domain tests still pass — confirming the shape logic now lives entirely in tensor. Suite 2408/2408.