llama-mmap: hint THP on mmap'd weights (Linux)

[Marxist-Leninist]

Issue madvise(MADV_HUGEPAGE) on the read-only file mapping used for model weights on Linux. For a 1 GB model this drops the potential page count from ~262K 4KB pages to ~512 2MB pages, reducing TLB pressure and (more importantly) reducing the number of re-faults when pages get evicted under memory pressure.

Linux-only, guarded by defined(MADV_HUGEPAGE) and __linux__. Skipped when numa is set. No-op where THP is disabled.

Bench on a Skylake-SP VM, Bonsai-8B Q1_0, -fa on -ctk q8_0 -ctv q8_0 -t 12 -ub 128: neutral (~9.5 t/s tg128 both before and after) because the VM isn't memory-constrained. The change is intended for systems where the mapping does get evicted under pressure (constrained laptops, containers).

Tried the same hint on ggml_aligned_malloc for KV/activation buffers as well — that showed a ~5% regression with no visible AnonHugePages, dropped that half.

[Marxist-Leninist]

Heads-up: the failing labeler check here is a pre-existing issue in the fork's workflow config, not related to this patch.

The PrismML-Eng labeler workflow (.github/workflows/labeler.yml) hard-codes repository: "ggml-org/llama.cpp" on its checkout step and reads the labeler config from there. Upstream ggml-org/llama.cpp/.github/labeler.yml still uses the v5 composition syntax:

server/webui:
    - changed-files:
        - all:
            - any-glob-to-any-file:
                - tools/server/webui/**

actions/labeler@v6 removed the all: / any: composition keys, so it errors with Unknown config options were under "changed-files": all on every PR — not just this one. The fix is either:

1. Bump actions/labeler back to v5 in the workflow, or
2. Flatten the server/webui entry in upstream's labeler.yml (drop the all: wrapper since it only contains one child):

server/webui:
    - changed-files:
        - any-glob-to-any-file:
            - tools/server/webui/**

Happy to open either fix as a separate PR if you'd like.