Add regular expand-convolute encoders

[ladnir]

Summary

• add a two-sided-regular binary expand-convolute encoder with degree 10/5 and memory 15 for Silent OT
• add a field-valued degree-26/13, memory-4 streaming encoder for GF(2^128) and Goldilocks Silent VOLE
• add a materialized reference encoder for the independently labeled finite-field paper ensemble
• add dimension adapters, end-to-end benchmarks, algebraic equivalence tests, rank checks, and Silent OT/VOLE integration tests

Application and proof boundary

Silent OT uses the binary profile because its choices are bits. Silent VOLE uses the field profile: its convolution regenerates full-field coefficients and performs ambient-field multiplications, so GF(2^128) components are mixed rather than encoded as 128 independent binary lanes. Silent VOLE explicitly rejects the binary RegularEc10x5x15 profile.

The binary degree-10/5, memory-15 ensemble has a finite distance certificate at k = 1,048,575 with failure probability below 2^-32.4990 for the stated cutoff.

The companion paper separately certifies the independently labeled field ensemble over GF(2^128) with degree 26/13 and memory 4. At n = 2,097,134 and distance cutoff 1,015,822, its sampling-failure probability is below 2^-135.20. The low-memory streaming implementation is deliberately marked heuristic: its compact seeded schedule and unit expander labels in characteristic two are not the independently labeled paper ensemble. Its full-field convolution still provides the field mixing needed by the VOLE implementation.

Neither profile becomes the default in this PR.

Validation

• RelWithDebInfo build with Silent OT and Silent VOLE enabled
• RegularEc encoder/configuration tests
• end-to-end GF(2^128) and Goldilocks Silent VOLE tests under regular and stationary noise
• dedicated Silent OT integration test
• every end-to-end VOLE benchmark verifies all output correlations after timing

At 2^20 GF(2^128) correlations, the field-valued RegularEc26x13x4 profile measured 0.792 s median (60.6 MiB/s), versus 0.210 s (228.5 MiB/s) for the current BlkAcc3x32 default. The new profile is therefore proof-motivated and opt-in, not proposed as a speed improvement.