Multithread the order-2 likelihood kernels on the RCPP backend

[manuhuth]

Summary

The RCPP backend was entirely single-threaded, and cocoReg's cores argument was dead — nothing read it. This PR parallelizes the two kernels that dominate order-2 fit time, likelihoodGP2 and likelihoodGP2cov, using C++11 std::thread (no OpenMP, no new dependencies), and makes cores live.

• likelihoodGP2: the per-(y,z) g-vector cache — rebuilt every optimizer iteration, the dominant cost — is now built concurrently across the distinct pairs, over read-only lookup tables ensured up front.
• likelihoodGP2cov: each observation's contribution is independent (lambda_t varies, nothing is cacheable), so per-t log-terms are computed concurrently with thread-private tables into a preallocated vector.
• In both, the final accumulation stays serial in the original order, so results are bit-identical for any thread count — verified against a pre-change build (identical() on NLL values at 1 and 14 threads, plus fitted parameters and log-likelihoods of full fits).
• The Poisson order-2 variants use the same kernels (eta = 0) and benefit equally. Order-1 kernels are microsecond-scale, where thread-spawn overhead would exceed any gain — they stay serial.
• cocoReg(cores=) defaults to physical cores − 1 (mirroring the JULIA_NUM_THREADS convention from Default JULIA_NUM_THREADS to half the physical cores for the Julia backend #9), capped at 2 while _R_CHECK_LIMIT_CORES_ is set so CRAN checks stay policy-compliant.
• Worker threads never touch the R API: the only stop() condition is checked on the main thread before spawning, and worker exceptions are re-raised as an R error after joining.
• Allocations stay minimal: one job list / one T-length vector per call plus per-thread tables that reuse capacity; dynamic scheduling via a single atomic counter.
• src/Makevars/Makevars.win add -pthread for the Linux/Windows toolchains.

Benchmarks

Measured on 14 physical cores (Apple Silicon), GP2 with T = 1000:

	1 core	13 cores	speedup
likelihoodGP2 (per call)	9.20 ms	1.43 ms	6.4x
likelihoodGP2cov (per call)	10.73 ms	1.27 ms	8.4x
cocoReg GP2 fit	6.30 s	1.09 s	5.8x
cocoReg GP2cov fit	11.44 s	1.50 s	7.6x

Test plan

• Bit-identity harness: baseline values (NLL, fitted parameters, log-likelihoods) saved from a pre-change worktree build, compared with identical() at cores = 1 and cores = 14 — all exact matches.
• devtools::test(): FAIL 0 / WARN 29 / PASS 18 — identical counts to the pre-change baseline.
• R CMD check: 0 errors, 0 warnings, 1 environment-only NOTE (future file timestamps).

🤖 Generated with Claude Code