Tune Linux native build flags and xxhash compilation

[yawkat]

Summary

• switch Linux JNI build from one-shot compile/link to per-source object compilation in pom.xml, with distro-like default Linux CFLAGS and a dedicated native.xxhash.cflags property
• compile xxhash.c with -fno-tree-vectorize by default on Linux and keep all flags overridable via -Dnative.*
• reduce workflow duplication by centralizing common Linux flags in job-level env vars and using per-arch deltas only in the matrix

Validation

• ./mvnw -q verify passed
• benchmarked bundled at.yawk.lz4:lz4-java:1.10.1 vs this branch's freshly built native lib (-Djava.library.path=target/classes/net/jpountz/util/linux/amd64) on pinned core:
  • STREAM32K (283 MiB TPCDS-like): 296.07 MiB/s -> 321.01 MiB/s (~+8.4%)
  • HASH32_BLOCK32K: 3391.17 MiB/s -> 6114.35 MiB/s (~+80.3%)

Notes

• perf guard benchmark intentionally not added per request

[yawkat]

Note: This is an AI-generated PR to attempt to fix the performance regression in #30. I asked the agent to adapt flags to better match distro lz4 builds, and to measure the performance impact.

[github-actions]

🚀 Preview deployed to https://7d0c5bcd.lz4-java-yawk-at.pages.dev

[yawkat]

Hmm still some issues, I forgot that the cross-platform build needs to be triggered manually