DeepSeek/GLM: fix reasoning wire shapes (disable toggle + effort tier)

[yogthos]

Two fixes, both empirically verified against the live hosted DeepSeek (api.deepseek.com) and GLM (open.bigmodel.cn) APIs.

1. Disable-thinking one-shots were a no-op on hosted DeepSeek + GLM

reasoning_disable_for_kind (provider/summarize.rs) turns off extended reasoning for the tool-less one-shots (summarizer / critic / approval). It sent {"chat_template_kwargs":{"thinking":false}} for deepseek/glm — but that's the vLLM/SGLang self-hosted convention. The hosted DeepSeek and GLM APIs silently ignore it and disable reasoning only via {"thinking":{"type":"disabled"}}.

Measured: with chat_template_kwargs, deepseek-v4 kept emitting ~130 reasoning tokens (GLM kept reasoning too); with thinking:{type:disabled}, both went to zero. So every summary/title/critic/approval call on hosted deepseek/glm was paying for reasoning it meant to skip (~2× latency).

Fix: send the thinking toggle for deepseek/glm; keep chat_template_kwargs for opencode/custom/openrouter (self-hosted backends).

2. DeepSeek reasoning-effort control was near-inert (wrong shape + capped tier)

build_provider_additional_params (agent_loop/rig_stream_factory.rs) sent deepseek a nested {"reasoning":{"effort":…}} and capped Xhigh→"high". Measured on a fixed reasoning prompt:

• top-level reasoning_effort: low ≈ 384 vs max ≈ 731 reasoning tokens — clean, non-overlapping ~2× separation.
• nested reasoning:{effort}: low and max means overlap → effectively ignored.

So dirge's effort control for DeepSeek was a near-no-op. Fix: give deepseek its own arm that sends top-level reasoning_effort and can reach the "max" tier (Xhigh), which OpenAI rejects. openai/glm/custom/openrouter keep the nested shape; GLM ignores effort entirely (fixed depth, measured), so it's unchanged.

Scope / categorization

• Keyboard navigation: Emacs-style bindings + kill ring #1 applies to DeepSeek + GLM (both hosted "thinking API" providers).
• Tool output display: visible results, colorized diffs, smart truncation #2 is DeepSeek-specific.
• A third idea (connection warmup) was investigated and dropped: the rig client is built once per session and reqwest already pools with a ~90s keepalive, so there's no keepalive gap; the only residual is a one-time first-handshake, not worth hand-building a TLS client for (dirge-fiis).

Tests: new unit tests for the deepseek effort mapping + top-level-vs-nested shape assertions; existing per-provider tests updated. fmt/clippy clean, full suite green (1 pre-existing flaky h7_glm smoke test unrelated).

Closes dirge-r9k8, dirge-f1su.