Exact FLOP costs for batched/broadcast contractions; unify on FMA=2

[spMohanty]

Summary

Routes the binary-contraction family — matmul, dot, inner, vecmat, matvec, vecdot —
through a single shared cost path built on the symmetry-aware einsum accumulation model, so per-op
FLOP counts are exact for all operand layouts (batched, broadcast, mixed-rank, 1-D-promoted),
not just the 2-D case. Previously these ops used per-op fallback formulas for non-2-D inputs that
did not account for all batch/broadcast axes; counting now derives from the operation's einsum
contraction structure, so it matches fnp.einsum(<equivalent subscripts>) exactly. The
contraction-order path search is unified onto the same FMA=2 convention used for billing,
and the legacy cost fallback is removed — one cost model end-to-end.

What changes

• New shared helper _einsum_routed_binary — builds the op's einsum subscripts and routes cost +
  output-symmetry inference through _resolve_cost_and_output_symmetry (the path matmul/dot 2-D
  already used). All six wrappers use it.
• vecmat/matvec/vecdot now count batch/broadcast axes exactly and bill in FMA=2.
• matmul/dot N-D & mixed-rank and inner N-D route through einsum instead of the
  a.size * b.size fallback.
• Path search (≥3-operand einsum) uses the FMA=2 accumulation cost — identical to billing — and
  the legacy FMA=1 fallback is removed.
• linalg.lstsq uses matmul_cost directly (its 2-D×1-D workaround is no longer needed).
• Docs: corrected stale "FMA=1" labels and dropped references to the removed fma_cost setting.

Gaming-resistance

Complements the existing no-gaming property (symmetric cost ≤ dense) with the dual guard: contraction
cost cannot be under-counted by re-expressing a matmul as a batched vector op. New parity tests
assert every op equals its einsum equivalent across batched/broadcast/mixed-rank shapes and that
cost scales with the batch dimension.

Breaking change

FLOP costs change for the affected ops (exactness + FMA=2 unification). Consumers that pin or budget
on absolute FLOP counts should re-baseline. A @ A symmetric behavior (symmetry-aware cost +
SymmetricTensor output) and all 2-D costs already routed through einsum are unchanged.

Out of scope

• tensordot partial-contraction symmetry path (keeps direct_product_groups).
• A @ A.T symmetry detection for non-symmetric A (write as einsum).
• outer/kron/vdot (no contracted axis; already exact).

Test plan

• New parity + batch-scaling tests for all six ops; A@A symmetry preserved.
• Multi-operand einsum path-selection regression (chosen paths + billed totals).
• Updated snapshots: test_cost_formula_vs_code.py, test_issue_69_cost_parity.py[lstsq],
  test_fma_unification.py, PathInfo snapshots.
• Full suite, ruff, pyright, coverage ≥90%, client-server-sync, registry-conformance.

Test-suite robustness (no product code change)

• tests/accumulation/test_deletion_safety.py now restores the package's lazy upstream
  __getattr__ shim after it imports the local _paths/_path_random submodules. Without this,
  those imports leaked into the package __dict__ and shadowed the shim for the rest of the
  process, causing the custom/random-optimizer tests in test_opt_einsum_paths.py to fail under
  serial (-n 0) execution (they passed under xdist, which distributes the tests across workers).