Fix #453: shape in tensor_zero identity; pin projector grouping

[petlenz]

Closes #453.

Part 1 — tensor_zero identity ignores shape (real, fixed). Equality compared hashes, and the hash covers only the node id, so every zero of every shape was one key:

	before	after
0{3,2} == 0{2,4}	true	false
set{0{3,2}, 0{2,4}, 0{3,4}}	size 1	size 3

operator== now also compares dim and rank; operator< keeps the hash as a fast reject and falls through to (dim, rank).

The hash is deliberately unchanged, following the precedent of #443/PR #458 for tensor symbols: the hash is a fast reject and an ordering key, not an identity, and changing it churns hash-driven print order (measured there: several print-pinned tests flip). Verified unchanged here: pow(x,2)*y*A print order, and both invariants — tensor_scalar_mul(const,T).hash == T.hash, tensor_pow(T,const).hash == T.hash. All zero folds still work: A+0, 0+A, A-0, 0-A → -A, 2*0 → 0{2}, trace(0), det(0), norm(0).

Part 2 — the issue's claim does not hold; no fix needed. tensor_projector_simplifier does bucket candidates by argument hash, but the combining step already re-checks deeply:

// hash groups may alias (hash(c*T)==hash(T)); require deep equality
if (!(entries[i].argument == entries[j].argument))
  continue;

That guard came from f92f94e ("Fix #340, #341: replace raw hash-equality identity checks with deep comparison"), which the audit that filed this issue missed — it read the map declaration at line 112 but not the guard 27 lines later. Confirmed by probe: vol(A) + dev(2A) stays 2*dev(A)+vol(A), while vol(A) + dev(A) folds to sym(A). Bucketing by hash loses nothing, since equal arguments necessarily hash equal.

Decision: keep the file, add the missing test. The issue offered "fix the keying or delete it"; the keying is already correct, so the only real choice was deletion. Against it: the class is described in DOCUMENTATION.md and docs/simplifier-coverage.md, and the tensor rule-contract work has an open decision to either wire it in as the driven optional pass or drop it as redundant — deleting now preempts that, in the irreversible direction. What the issue actually worried about ("whoever wires it inherits the bug") is addressed by ProjectorPassGroupsOnlyDeepEqualArguments, which is the first test to exercise this class at all: it asserts the aliasing still holds, then that a mixed-argument sum does not combine. It passes on main by design — a lock-in, not a bug demonstration, and the PR says so rather than implying it caught something.

Singletons checked as the issue asked: scalar_zero, scalar_one, tensor_to_scalar_zero and tensor_to_scalar_one all have default constructors and carry no dim/rank or other per-instance state (the t2s constants pre-insert assumptions, identical for every instance), so identity by node id alone remains correct for them.

Tests: HashIdentitySweep.TensorZeroShapeIsPartOfIdentity (equality, ordering totality, a 3-entry map) and HashIdentitySweep.ProjectorPassGroupsOnlyDeepEqualArguments.

Negative control (tensor_zero.h restored from main with git show, rebuilt): the zero test fails on 7 assertions and passes with the fix; the projector lock-in passes either way as intended. Full suite 2407/2407 (gcc-14 Debug), clang-format clean.