feat: add file-backed constraint system loading

[sergeytimoshin]

This PR adds file-backed loading for serialized constraint systems:

• Adds generated ReadFromFile(path) support to curve-typed constraint systems.
• Uses mmap on supported platforms to decode directly from file-backed pages.
• Falls back to the existing ReadFrom path on unsupported platforms.
• Adds public helpers:
  • groth16.ReadCSFromFile(curveID, path) for R1CS.
  • plonk.ReadCSFromFile(curveID, path) for SparseR1CS.
• Keeps the existing serialization format unchanged.
• Hardens malformed serialized-data handling in shared constraint-system decoding, integer-compression decoding, and generated coefficient-table decoding so invalid lengths return errors instead of panicking or attempting huge allocations.

Why

Today ReadFrom reads the whole serialized constraint-system payload into a heap []byte before decoding it. For large circuits this creates a temporary memory spike: the process holds both the serialized payload copy and the decoded constraint-system structures during load.

When the source is already a file, this PR removes that intermediate heap copy on mmap-capable platforms. The decoded R1CS/SparseR1CS still owns normal Go heap structures after loading, so the mmap can be closed before returning. This is intentionally a load-time peak-memory reduction, not a long-lived mmap-backed constraint-system representation.

This is useful independently of mmap-backed Groth16 proving keys: users who persist compiled R1CS or SparseR1CS artifacts can reduce load-time memory pressure even when using the existing proving-key formats.

API Notes

The PR does not add methods to constraint.ConstraintSystem. That avoids breaking external implementations of the interface. Instead, the fast file path is exposed through backend-level helpers next to the existing NewCS constructors, and the generated concrete types also get ReadFromFile.

Benchmark

Local benchmark on Apple M4 Max, darwin/arm64:

BenchmarkReadCSFromFile/ReadFrom-16          196069 ns/op   588948 B/op   4271 allocs/op
BenchmarkReadCSFromFile/ReadFrom-16          170360 ns/op   588941 B/op   4270 allocs/op
BenchmarkReadCSFromFile/ReadFrom-16          173465 ns/op   588968 B/op   4271 allocs/op

BenchmarkReadCSFromFile/ReadCSFromFile-16    160856 ns/op   425195 B/op   4272 allocs/op
BenchmarkReadCSFromFile/ReadCSFromFile-16    158764 ns/op   425179 B/op   4272 allocs/op
BenchmarkReadCSFromFile/ReadCSFromFile-16    162448 ns/op   425194 B/op   4272 allocs/op

The benchmark shows the expected reduction in heap bytes from avoiding the serialized-payload copy. Runtime is comparable; allocation count has two extra small allocations from file mapping.

Validation

Passed:

go generate ./internal/generator/backend
go test ./constraint/...
go test ./internal/backend/ioutils ./backend/groth16 ./backend/plonk
go test ./backend/groth16 -run '^$' -bench BenchmarkReadCSFromFile -benchmem -count=3
git diff --check

Checklist Notes

• Documentation: public API comments were added on ReadCSFromFile and generated ReadFromFile; the PR body explains intended usage and limitations.
• Generated files: this PR intentionally updates generated constraint-system marshal/test files and their templates.