perf(zkatdlog): reduce CPU time and allocations in CSP range-proof crypto

[adecaro]

Fixes #1994

Summary

Optimizes the Compressed Sigma Protocol (CSP) range-proof package
(crypto/rp/csp) with no change to its public API, wire format, or verification
semantics:

• crypto/rp/csp/msm.go (new): dispatches small, fixed-size multi-scalar
  multiplications to direct Mul/Mul2 instead of gnark-crypto's
  MultiScalarMul. CSP's folding rounds shrink every vector down to length 2
  and then 1 in their last two rounds, and several call sites always operate
  on exactly 2 generators/scalars; MultiScalarMul's goroutine fan-out (sized
  for runtime.NumCPU()*2) costs more than it saves at n<=2. Benchmarked on
  BLS12-381 (M1 Max): direct Mul is ~2.5x faster than MultiScalarMul at
  n=1, Mul2 ~25-29% faster at n=2; from n=3 the two are at parity or
  MultiScalarMul wins, so smallMSM only special-cases n=1/n=2 and falls
  through to MultiScalarMul otherwise (strictly non-regressive). Applied at
  all 6 production MultiScalarMul call sites in csp.go/rp.go.

• crypto/math/native.go: adds NativeBatchInverseInto, a batch
  field-inversion helper that writes into caller-supplied buffers instead of
  allocating fresh ones on every call.

• crypto/rp/csp/lagrange_native.go: interpolateNative's outer loop
  previously allocated 5 fresh same-size slices/scalars per iteration.
  Rewritten to hoist every per-x buffer and scalar scratch outside the outer
  loop via NativeBatchInverseInto, so each is allocated once per call
  instead of once per iteration.

• crypto/rp/csp/transcript.go: Transcript.Absorb is called many times per
  proving/verification round; reuses a scratch buffer across calls instead of
  growing a fresh slice on every Absorb.

Benchmarked on BenchmarkBFProver (n=32/n=64, BLS12_381_BBS_GURVY, count=8,
benchstat): combined effect of the MSM dispatch and batch-inversion changes
is roughly -6% ns/op and -35-40% allocs/op (geomean), each individually
significant at p<0.01. No public API, wire format, or verification behavior
changed.

Test plan

• make checks
• make lint-auto-fix
• make unit-tests
• Targeted: go test ./token/core/zkatdlog/nogh/v1/crypto/rp/csp/... ./token/core/zkatdlog/nogh/v1/crypto/math/... ./token/core/zkatdlog/nogh/v1/crypto/rp/bulletproof/...
• benchstat before/after on BenchmarkBFProver, count=8