Improve CUDA 13.1 compatibility with newer glibc

[remer]

Summary

Make Krasis's Linux CUDA builds tolerate the CUDA 13.0/13.1 declaration
collision with newer glibc without patching the installed CUDA toolkit or
changing the system compiler.

• pass -U_GNU_SOURCE -D_DEFAULT_SOURCE through the shared build.rs NVCC
  argument path;
• apply the same contract to the independently compiled Marlin and
  FlashAttention package sidecars;
• force-include a narrowly scoped compatibility header which disables the
  optional libstdc++ pthread clock-wait paths made unavailable when
  _GNU_SOURCE is hidden; and
• include that header in the sidecar input hash so cached bundles cannot mask a
  compatibility-source change.

Windows behavior is unchanged. An explicitly selected KRASIS_NVCC_CCBIN is
still preserved.

Observed failures

With CUDA 13.1 and glibc 2.43, normal CUDA compilation failed with errors
equivalent to:

bits/mathcalls.h: error: exception specification is incompatible with that of previous function "rsqrt"
crt/math_functions.h: previous declaration of "rsqrt"

After the original compatibility flags were exercised through the release
sidecar builder, FlashAttention exposed the corresponding libstdc++ half of the
contract: its optional timed-mutex paths referenced GNU pthread clock APIs that
were intentionally no longer declared. The compatibility header handles both
sides in the affected translation unit.

RC10 validation (2026-09-02)

• Base: v1.0.21-rc.10
  (716bf316fbc001c1f61ad607362e05c1cb323c50).
• Three focused host-argument regression tests passed.
• git diff --check and Rust formatting checks passed.
• A complete Marlin regular + MoE sidecar build passed.
• The complete packaged FlashAttention BF16 and BF16-query/FP8-KV matrix passed
  for head dimensions 64, 96, 128, 192, and 256, causal and non-causal.
• Marlin and FlashAttention linked successfully; required exported symbols,
  manifest input hashes, artifact hashes, and build IDs all passed the
  builder's post-build contract.
• The resulting sidecar bundle was restored from its content-derived cache key
  and revalidated, proving the new header participates in cache identity.
• A completely fresh-target RC10 integration wheel completed against Python
  3.13; both packaged sidecars loaded from that wheel and 29 wheel-backed
  DeepSeek Vision/model-configuration tests passed.

The DeepSeek-only canary does not require Krasis's separate Flash Linear
Attention sidecars, so this validation does not claim a full multi-model
release-wheel matrix.

Compatibility scope

The observed CUDA 13.1/new-glibc combination is outside CUDA 13.1's validated
distribution matrix, so this is a portability improvement rather than a claim
that Krasis is defective on a supported CUDA configuration. NVIDIA's
cuda-samples PR #403 documents the same feature-macro workaround:
NVIDIA/cuda-samples#403

No credentials, account names, hostnames, network addresses, private paths, or
raw request/generated content are included in this change or PR body.

RC11 alignment — 2026-09-08

Merged upstream v1.0.21-rc.11 (9d25fb3a490499f1be78952079d361acfdb5c05c) without rewriting this
branch's history. The feature diff is unchanged; RC11 contributes its 13
context-default/launcher/manager/version/test files. Previous head preserved
at remer/krasis:preserve/pre-rc11-20260908/pr31.

The combined integration/rc11-ai-box tree
d8656f9577a7d469a7b0e2e987b61b14e258e1d5 builds a fresh release wheel.
Its targeted Rust suites pass: D-Spark 30/30, chat template 16/16, server 31/31.
Python model/configuration/vision/cache-gate contracts pass. One launcher parser
case assumes two visible GPUs; it passes with an isolated two-GPU inventory
fixture (parser-only, not a claim of multi-GPU runtime testing).

Live RC11 Vision qualification is pending. Historical RC6/RC10 throughput and
quality measurements are not RC11 Vision results. Session-cache operational
and D-Spark promotion gates remain live-test obligations.