llms/ollama: WithThink has no effect — think is serialized inside "options" (API expects top-level) and omitempty drops false

Two independent defects in the Ollama integration make WithThink (and, on current main, the ThinkingMode-derived Think flag) a silent no-op.

1. Wrong placement: think is sent inside options, but the Ollama API defines it as a top-level field.

Think is declared in the runner-Options struct (types.go#L170 @v0.1.14, unchanged on main):

PenalizeNewline  bool    `json:"penalize_newline,omitempty"`
Think            bool    `json:"think,omitempty"` // Ollama 0.9.0+ reasoning mode

ChatRequest serializes that struct under "options" and has no top-level think field. The Ollama API expects think at the top level of both /api/chat and /api/generate (docs/api.md: "think: (for thinking models) should the model think before responding? Can be a boolean or a thinking level"). Unknown keys inside options are ignored by the server.

The repo's own recorded fixture shows the resulting wire format (llms/ollama/testdata/TestWithThink.httprr):

{"model":"gemma3:1b","messages":[...],"format":"","options":{"temperature":0,"think":true}}

2. omitempty on a plain bool: WithThink(false) serializes to nothing at all.

false is the zero value, so even with the placement fixed, "explicitly disable thinking" could never be expressed. For thinking models (qwen3, deepseek-r1, …) the server default is to think — false is precisely the value users need to send.

Net effect: WithThink(true) → ignored key inside options; WithThink(false) → omitted entirely; both silent. This likely also explains #1460 — the request-side flag never reaches the server, so all thinking-related options appear dead.

Measured impact: paperless-gpt exposes OLLAMA_THINK backed by WithThink; the flag has no effect. With qwen3:14b our metadata pipeline ran at ~3 min/document (model thinking) vs ~8 s/document once think: false actually reached the server top-level (injected via a proxy). In a paired n=23 run (strict-JSON classification, qwen3:14b), disabling thinking was equal or better on accuracy (87.0 % vs 82.6 % on the main field).

Suggested fix — move Think out of Options into ChatRequest (and the generate request) as a pointer, so unset ≠ false, and populate it from the existing option paths (WithThink, the ThinkingMode mapping in ollamallm.go):

type ChatRequest struct {
	Model     string     `json:"model"`
	Messages  []*Message `json:"messages"`
	Stream    bool       `json:"stream,omitempty"`
	Format    string     `json:"format"`
	KeepAlive string     `json:"keep_alive,omitempty"`
	Think     *bool      `json:"think,omitempty"` // top-level per Ollama API; pointer keeps unset ≠ false
	Options   Options    `json:"options"`
}

(A follow-up could widen the type to bool | "low" | "medium" | "high" per current Ollama semantics.)

Affected: v0.1.14 and current main.