feat: incremental AST cache to eliminate redundant AstEncoder work

[daniplatform]

Summary

Adds an incremental AST cache that eliminates redundant AstEncoder work the pure-Python json.dumps(ast_tree, cls=AstEncoder) step, which is O(N nodes) and dominates parsing cost (ast.parse itself is C and negligible).

Three-level hierarchy:

• L1 in-memory mtime guard: zero work on a hit within a single process run.
• L2 disk content-hash guard: no parse/encode across runs.
• L3 chunk-aware per-function / per-class subtree reuse when a file only partially changes.

Changes

• src/pyspector/ast_cache.py — cache implementation. Persistence is JSON + base64 + zlib, deliberately not pickle (pickle executes arbitrary code on load, unsafe for cache files living in an untrusted repo directory). Atomic writes, LRU eviction, graceful degradation to L1-only when the cache dir can't be created.
• src/pyspector/_ast_encode.py — shared AST → JSON encoder extracted as the single source of truth for the schema consumed by the Rust core, eliminating encoder drift between cli.py and the cache.
• src/pyspector/cli.py — wires the cache into get_python_file_asts via a new optional cache parameter, and re-exports AstEncoder from _ast_encode (so from pyspector.cli import AstEncoder keeps working). The cache is bypassed when enable_syntax_warnings is True, preserving that diagnostic (the cache suppresses SyntaxWarning internally).
• tests/unit/ast_cache_test.py — 51 unit tests (all green).

Compatibility

This branch was rebased onto the latest main, so it integrates cleanly on top of the recent CLI work (--stats, --debug, exclude pre-pass, syntax-warning handling, absolute file_path). All of those upstream behaviours are preserved; the cache only adds an optional fast path. The output AST JSON is byte-for-byte identical to the non-cached path.

Testing

• tests/unit/ast_cache_test.py: 51 passed.
• AST JSON produced via the cache verified identical to the direct ast.parse + AstEncoder path.
• Pre-existing failures unrelated to this change (stale compiled _rust_core vs. new rules-TOML schema, missing bs4, and an existing absolute-vs-relative file_path assertion in test_get_asts.py) reproduce identically on main and are out of scope here.

[ParzivalHack]

Hey @daniplatform, great idea on reducing the workload of AstEncoder, and in my tests this PR also increases PySpector's average scanning speed by 41.90%, so really a great addition. I require no edits. Merging :D