mlkem: reduce allocations

[qmuntal]

Most of MLKEM allocations can be avoided by using big-enough buffer sizes when encoding/decoding CNG blobs. We know those max sizes, so use.

goos: windows
goarch: arm64
pkg: github.com/microsoft/go-crypto-winnative/cng
                        │   old.txt    │              new.txt               │
                        │    sec/op    │   sec/op     vs base               │
MLKEMKeyGen-12            91.96µ ± 16%   81.21µ ± 7%  -11.69% (p=0.001 n=7)
MLKEMEncaps-12            32.45µ ± 16%   28.75µ ± 2%  -11.39% (p=0.001 n=7)
MLKEMDecaps-12            45.30µ ±  1%   44.02µ ± 2%   -2.84% (p=0.002 n=7)
MLKEMRoundTrip/Alice-12   132.7µ ±  3%   124.8µ ± 0%   -5.98% (p=0.001 n=7)
MLKEMRoundTrip/Bob-12     29.63µ ±  3%   28.70µ ± 0%   -3.14% (p=0.001 n=7)
geomean                   55.61µ         51.66µ        -7.09%

                        │   old.txt    │                 new.txt                  │
                        │     B/op     │     B/op      vs base                    │
MLKEMKeyGen-12            1.438Ki ± 0%   0.000Ki ± 0%  -100.00% (p=0.001 n=7)
MLKEMEncaps-12            3.656Ki ± 0%   2.406Ki ± 0%   -34.19% (p=0.001 n=7)
MLKEMDecaps-12             128.00 ± 0%     32.00 ± 0%   -75.00% (p=0.001 n=7)
MLKEMRoundTrip/Alice-12   1600.00 ± 0%     32.00 ± 0%   -98.00% (p=0.001 n=7)
MLKEMRoundTrip/Bob-12     3.656Ki ± 0%   2.406Ki ± 0%   -34.19% (p=0.001 n=7)
geomean                   1.303Ki                      ?                      ¹ ²
¹ summaries must be >0 to compute geomean
² ratios must be >0 to compute geomean

                        │  old.txt   │                new.txt                 │
                        │ allocs/op  │ allocs/op   vs base                    │
MLKEMKeyGen-12            3.000 ± 0%   0.000 ± 0%  -100.00% (p=0.001 n=7)
MLKEMEncaps-12            4.000 ± 0%   3.000 ± 0%   -25.00% (p=0.001 n=7)
MLKEMDecaps-12            2.000 ± 0%   1.000 ± 0%   -50.00% (p=0.001 n=7)
MLKEMRoundTrip/Alice-12   5.000 ± 0%   1.000 ± 0%   -80.00% (p=0.001 n=7)
MLKEMRoundTrip/Bob-12     4.000 ± 0%   3.000 ± 0%   -25.00% (p=0.001 n=7)
geomean                   3.438                    ?                      ¹ ²
¹ summaries must be >0 to compute geomean
² ratios must be >0 to compute geomean

[Copilot]

Pull Request Overview

This PR optimizes ML-KEM operations by reducing memory allocations through pre-allocated buffers of known maximum sizes. The main change is to allocate buffers upfront using the maximum size needed (ML-KEM 1024 sizes) instead of making two allocation calls per operation.

Key changes:

• Pre-allocate buffers using ML-KEM 1024 sizes to avoid allocation overhead for both 768 and 1024 parameter sets
• Simplify error handling by converting helper functions from returning errors to panicking (errors are still properly propagated from public APIs)
• Remove unnecessary runtime.KeepAlive calls and optimize function signatures

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
cng/mlkem.go	Core optimization changes: added buffer size constants, refactored blob creation functions to use pre-allocated buffers, simplified error handling in internal helpers
cng/mlkem_test.go	Added t.Parallel() calls to test functions to enable parallel test execution

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.