feat(prefetch): add file selection (allow_patterns/ignore_patterns) to hf_model

[mhotan]

Intent

flyte.prefetch.hf_model prefetches the entire HuggingFace repo. That makes it unusable for GGUF repos, which ship many quantizations at one commit (100s of GB) when a caller wants exactly one (~18 GB). This adds file selection so hf_model can prefetch just the files you want — the missing piece for using the prefetch → model-artifact pattern with flyteplugins-llamacpp (#1521).

What changed

• allow_patterns / ignore_patterns threaded through HuggingFaceModelInfo → hf_model() → store_hf_model_task into both download paths.
• Streaming path (_stream_to_remote_dir) now lists the repo recursively (list_repo_files) and filters with huggingface_hub.utils.filter_repo_objects, preserving each file's repo-relative subpath — so nested quant layouts are fully captured, not just top-level files. (Previously a flat ls silently skipped subdirectories.)
• Snapshot-download fallback forwards the patterns; the unfiltered call is byte-for-byte unchanged.
• Artifact metadata records the selected patterns (allow_patterns/ignore_patterns attrs), so a stored subset (e.g. one quant) is identifiable beyond the source commit.
• CLI: flyte prefetch hf-model gains --allow-pattern / --ignore-pattern (repeatable).
• Sharding intentionally ignores the patterns (it needs the full weights).
• Examples: new examples/genai/llamacpp/ — the GGUF counterpart to vllm/sglang: prefetch one quant (allow_patterns=["*q4_k_m*"]) → model artifact → serve via the llama.cpp plugin, bound by ArtifactValue.

Backward compatible: default (None) prefetches the whole repo exactly as before — existing callers and the exact-kwargs snapshot test are unaffected.

How it was tested

• pytest tests/flyte/prefetch/test_hf_model.py — 65 passed, 1 skipped. New tests cover: pattern fields default None + json round-trip; streaming filters to matching files (incl. nested) and preserves subpaths; snapshot fallback forwards patterns (and omits the kwargs when unset); metadata records the patterns; hf_model threads patterns into the info.
• pytest tests/cli/test_prefetch.py — 16 passed.
• ruff check + ruff format --check clean on all changed files.

Notes

First of a two-PR stack; a follow-up adds a lazy object-store FUSE serving mode for GGUF artifacts (the scale-to-zero-clean delivery that download/stream_model don't cover). This PR stands alone.