Fix #475 and #476: odd-power sqrt fold and permute index validation

[petlenz]

Closes #475.
Closes #476.

Base is fix-336-pow-pow-domain (PR #462), not main — #475 relaxes the gate that PR adds, so this merges after it.

#475 — pow(sqrt(x), n) folds for an odd n. An odd n leaves a denominator of 2 in n/2, so pow(x, n/2) is non-real for x < 0 exactly where sqrt(x) is; the two sides agree for every real x. An even n would make the right side finite there, so it still requires a nonnegative radicand.

Expression	Before	Now
pow(sqrt(x), 3), unassumed x	pow(sqrt(x),3)	pow(x,3/2)
pow(sqrt(x), 2), unassumed x	pow(sqrt(x),2)	unchanged — correctly blocked
pow(sqrt(x), -3)	unfolded	pow(x,-3/2)

Numerically verified at x = −4, −0.5, 0, 0.5, 4 for both parities, and for n = -3 / n = -2: folded results match std::pow(std::sqrt(x), n) everywhere, including NaN for negative x and inf at x = 0.

Spelling variants (3, 3.0, 6/2) all fold and evaluate identically. Note the resulting exponent's representation still differs (3/2 vs 1.5) on this base, so the test compares by value; #459/#473 are what make numerically equal spellings one value.

Test change, called out explicitly. Scalar_PowSqrtSimplification asserted pow(sqrt(x), _3) → "pow(sqrt(x),3)". That assertion was introduced by #462's blanket gate; on main it reads "pow(x,3/2)". This restores the original for the odd case and keeps #462's new even-case assertion, with a comment stating the parity condition.

#476 — permutation validated in the node constructor. Only permute_indices() checked that the index sequence covers the operand's rank, so a direct construction could build a node claiming to be a rank-2 transpose while wrapping a rank-4 tensor, which is_trans_of accepted. The check moved into a shared validate_permutation helper that the constructor and the factory both call, so rebuild, substitution and direct construction are covered — the same precedent #438 set for the inner/outer product wrappers. The factory's duplicated loop is gone and its messages are unchanged.

Internal direct constructions audited, all pass validated sequences: tensor_rebuild_visitor.h:112 (reuses the node's own indices), the composition fold and the factory tail in tensor_functions.h, and trans() (rank-2 gated, {2,1}).

Since the malformed node can no longer be constructed, "is_trans_of cannot be fooled" is now enforced at construction rather than asserted separately.

Tests: ScalarEval.PowOfSqrtFoldsForOddExponents and TensorShapeValidation.PermuteWrapperCtorValidatesIndices (mismatched rank, out-of-range, duplicate, plus valid rank-2 and rank-4 permutations still constructing).

Negative control (both production files restored from the base with git show, rebuilt clean): all three of PowOfSqrtFoldsForOddExponents, PermuteWrapperCtorValidatesIndices and the updated Scalar_PowSqrtSimplification fail; all pass with the fixes.

Full suite 2414/2414 (gcc-14 Debug) and 2600/2600 parser-enabled; clang-format clean.