perf(cli): stop importing sentence_transformers for every command

[r0h1tb]

Found while running a full UAT sweep of the CLI against a live index — every command had a hard ~2.9s floor, including pure graph lookups.

Stacked on #50 and #56. Review those first; this branch contains them.

Cause

import ast_rag.cli costs 3.65s, and -X importtime attributes 3.05s of it to sentence_transformers (and transitively torch/transformers), reached via services/__init__ → embedding_manager:

    3651 ms   ast_rag.cli
    3508 ms     ast_rag.services.parsing.parser_manager
    3462 ms           ast_rag.services.embedding_manager
    3054 ms             sentence_transformers
    1447 ms                   transformers.configuration_utils

Only semantic search needs the model. goto, callers, refs, sig, blocks, cache-stats are pure Neo4j lookups and never touch it — but they all paid for it.

Fix

Move the import into _get_model(), where it's actually used, with TYPE_CHECKING for the annotations. Same shape as the watchdog fix in #50.

import ast_rag.cli   3.65s -> 0.61s

Measured end to end

15 CLI scenarios against a live index (raged indexing itself — 965 nodes, 3,782 edges):

	before	after
median command latency	3.068s	0.558s
fastest command	2.864s	0.487s
goto	3.371s	0.655s
callers	2.887s	0.487s

query still pays the model load on first use, as it must — it's unchanged at ~12s cold, and it still returns its full result set, so the lazy path is exercised and works.

For an interactive tool, and especially for an agent calling goto/callers in a loop, a 5.5x cut on the common path is the difference between usable and not.

Verification

pytest tests/    3 failed, 191 passed, 1 xfailed
ruff check       All checks passed!

Baseline on main is 3 failed, 174 passed — same three pre-existing failures (fixed in #51). No test changes here; the existing suite covers that embeddings still work, and the UAT sweep confirms query behaves identically.

[r0h1tb]

Rebased onto the rebased #56 (this PR is stacked on it, so #56 needs to merge first — until then the diff here shows #56's commits too).

Resolved one conflict in embedding_manager.py: #56 added Callable to the typing import while this branch added TYPE_CHECKING. Both are used, so the import is now from typing import TYPE_CHECKING, Callable, Optional.

Measured effect of this PR — importing ast_rag.services on each branch:

branch	pulls in sentence_transformers?	import time
main	yes	5.80s
#56	yes	4.78s
this branch	no	0.48s

That's ~12x on every CLI invocation, including --help, because services/__init__.py imported embedding_manager eagerly and that imports sentence_transformers at module scope.

Incidental confirmation that this is worth having: while setting up to verify these PRs I couldn't run the parser tests without installing torch + sentence-transformers, purely because of this import chain. On this branch they run without the ML stack present.

Verified against main (6f71a9b): ruff check and ruff format --check clean; pytest tests/ gives 216 passed, 1 skipped, 1 xfailed vs. main's 209/1/1 — the +7 are #56's tests carried in the stack, nothing pre-existing changed state.