client: widen sync generate think type to accept string levels

[anxkhn]

The think parameter accepts either a bool or the 'low'/'medium'/'high'
string levels, and the request model serializes all of them
(GenerateRequest.think in ollama/_types.py is
Optional[Union[bool, Literal['low', 'medium', 'high']]]).

PR #553 (commit aa4b476, "add support for high/medium/low think values")
widened this annotation across Client.chat, AsyncClient.generate, and
AsyncClient.chat, but missed the sync Client.generate overloads and
implementation, which stayed at Optional[bool].

As a result, calling the documented string levels through the sync generate path,
e.g. ollama.generate(model=..., prompt=..., think='high') (the module-level
generate is bound to the sync Client.generate), is flagged as a false type
error by static type checkers, even though it works at runtime and the equivalent
chat() and async generate() calls type-check cleanly.

What this changes

• Widen the three sync Client.generate occurrences (the two @overload stubs
  and the implementation) from think: Optional[bool] = None to
  think: Optional[Union[bool, Literal['low', 'medium', 'high']]] = None, so
  they match Client.chat and AsyncClient.generate exactly. Union, Literal,
  and Optional are already imported; no new imports.
• Add a signature-parity test (test_generate_think_annotation_matches_chat)
  asserting that Client.generate, Client.chat, and AsyncClient.generate
  share the same think annotation, guarding against this kind of drift in the
  future.

This is a pure type-annotation change: no runtime behavior changes, and the blast
radius is limited to static type checking of Client.generate / the module-level
generate.

Reproduction (before this PR)

import ollama

# Documented string think levels, flagged as a false type error on the sync path:
ollama.generate(model='gpt-oss', prompt='hi', think='high')

A type checker reports, on main:

error: Argument of type "Literal['high']" cannot be assigned to parameter "think"
of type "bool | None" in function "generate"
error: No overloads for "generate" match the provided arguments

The same call through chat() or async generate() type-checks cleanly today;
after this PR the sync generate() call does too.

Testing

• uv run pytest ollama tests -q -> all tests pass (the CI scope; examples/
  is excluded as it needs a live server).
• uvx ruff check . and uvx hatch fmt --check -> clean.
• New test is a real regression guard: it fails on the current annotation
  (Optional[bool]) and passes after the widening.

[vaderollama]

LGTM — fixes the annotation parity gap where sync Client.generate was left with Optional[bool] while Client.chat and AsyncClient.generate already accepted Union[bool, Literal['low','medium','high']]. Good regression test checking annotation equality across all three methods.

@ParthSareen — ready to merge.