Add generalized semiring reductions to einsum VM and JIT

[imlvts]

Generalize the einsum contraction to an arbitrary commutative semiring: a Reduce operator {Sum, Prod, Max, Min} folding contributions into the output, over a Combine operator {Mul, Add, Min, Max} merging the input operands — so max-product (A[i,j,k] = max_q B[j]*C[i,k,q]*D[k,q]) and min-plus (tropical / shortest-path) contractions run through the same machinery as plain sum-product einsum. The design is modelled after the Julia library Tullio.jl (noted as a footnote in the module docs).

• New Scalar element trait (tensor.rs) carrying the per-operator identity table: Sum→0, Prod→1, Max→-inf/T::MIN, Min→+inf/T::MAX. Empty reductions yield the identity.
• VM: compile_reduce / einsum_reduce / einsum_reduce_dyn. The einsum_reduce wrappers pre-fill outputs with the reduce identity (overwrite semantics); compile_reduce + Program::exec keeps accumulate semantics.
• Sparse row iteration (structural-zero skipping) is only sound when a missing operand annihilates its contribution to the reduce identity — precisely (Sum, Mul). Other op pairs iterate sparse inputs densely in the VM; the JIT rejects them with Unsupported.
• JIT: EinsumF32Jit::compile_reduce / einsum_jit_reduce, swapping fmax/fmin/fmul/fadd into the existing register-accumulator and RMW paths. (Sum, Mul) emits IR identical to before.
• Tests: semiring unit tests (incl. integer MIN-identity, empty ranges, AccFlush RMW edge, sparse-with-max dense fallback) plus a 481k-case differential sweep (tests/reduce_sweep.rs) comparing VM and JIT against a naive reference over all specs on {a,b,c}, 7 op pairs, all sparse masks — bit-exact. Existing 19.5M-case einsum_sweep passes unchanged.

VM (interpreted) einsum regressed ~5–15%:

Benchmark (VM einsum, base 5464713 → head f8b6fe0)	base	head	Δ
Dense matmul 16×16, einsum_homogenous (concrete)	65.6 µs	69.5 µs	+6.0%
Dense matmul 64×64, einsum_homogenous (concrete)	3.72 ms	3.97 ms	+6.8%
Dense matmul 128×128, einsum_homogenous (concrete)	29.3 ms	31.3 ms	+6.7%
Dense matmul 256×256, einsum_homogenous (concrete)	232.5 ms	247.5 ms	+6.5%
Dense matmul 16×16, einsum (&dyn)	77.4 µs	86.7 µs	+12.1%
Dense matmul 64×64, einsum (&dyn)	4.46 ms	5.05 ms	+13.3%
Dense matmul 128×128, einsum (&dyn)	35.1 ms	41.0 ms	+17.0%
Dense matmul 256×256, einsum (&dyn)	279.0 ms	317.2 ms	+13.7%
Dense matmul 128×128, einsum_homogenous (Tensor)	36.3 ms	38.5 ms	+6.2%
Dense matmul 256×256, einsum_homogenous (Tensor)	276.8 ms	303.0 ms	+9.5%
CSR×CSR side=5, einsum_homogenous (Csr)	47.7 µs	52.3 µs	+9.5%
CSR×CSR side=10, einsum_homogenous (Csr)	518.2 µs	561.0 µs	+8.3%
CSR×CSR side=20, einsum_homogenous (Csr)	56.0 ms	57.1 ms	+2.0%
CSR×Dense 1000×1000 × 1000×8, einsum (&dyn)	576.9 µs	636.5 µs	+10.3%
CSR×Dense 8000×8000 × 8000×16, einsum (&dyn)	7.93 ms	8.84 ms	+11.6%
CSR×Dense 8000×8000 × 8000×64, einsum (&dyn)	26.1 ms	29.5 ms	+13.1%
JIT matmul 128×128 (unaffected)	2.26 ms	2.30 ms	+1.6%
JIT matmul 256×256 (unaffected)	16.35 ms	16.48 ms	+0.8%

[MesTTo]

the VM regression looks like per-element dispatch: Reduce::apply and Combine::apply match on the enum, and the inner loop reads self.combine/self.reduce and calls them per element. the JIT resolves the op at codegen so it's fine, the VM pays the match every element.

tried lifting (Reduce, Combine) to type params and branching once in exec, +133/-18 in einsum.rs only. by instruction count the delta is -6.84%, both arms at 0.00% variance:

jit_bench, same fixed workload, instructions retired:
  this PR:  ~2.030 trillion
  +patch:   ~1.891 trillion   (-6.84%)

jit_bench.rs is byte-identical across both builds so the whole delta is the einsum change, and the JIT path is in that total too so the VM-only drop is a bit bigger. correctness holds with it applied: the linalg suite, your 481,278-case reduce sweep, and the 19,465,300-case exhaustive sweep all pass.

two caveats: i only measured (Sum, Mul), jit_bench doesn't run min-plus or max-product so i haven't checked those. and the extra instantiations cost size, ~+4.8% here and more on the rlib.

patch's here if it's useful: https://gist.github.com/MesTTo/0d470c97bfecfe503f79e026a537abe5

[MesTTo]

accidental comment