[Bug]: Reasoning effort sent to models that do not support it

[tivris]

Bug Description

Setting reasoning.enabled = false and effort = "none" in .forge.toml does not prevent the effort parameter from being sent to the Anthropic API. Models that do not support output_config.effort (e.g. Claude Haiku 4.5) reject the request with a 400.

Reported in #2924 by @pcastellazzi.

Root Cause

Two interacting bugs:

1. Config merge direction (crates/forge_app/src/agent.rs:164-167): The merge uses overwrite_none with the agent as base, so the forge agent's reasoning.enabled: true (hardcoded in forge.md:5-6) silently overrides the user's enabled: false. The user config can never turn off reasoning through .forge.toml.

2. Effort::None semantics (crates/forge_app/src/dto/anthropic/request.rs:153): Effort::None is documented as "No reasoning; skips the thinking step entirely" but the Anthropic DTO maps it to OutputEffort::Low, which enables low-effort reasoning. The OpenAI path already handles this correctly (maps to "none").

Fix

1. Reverse the merge direction so user config takes priority (matching the compact config pattern already used in the same file).
2. When Effort::None reaches the Anthropic DTO, return (None, None) instead of OutputEffort::Low.

[tivris]

Closing as duplicate of #2924. The fix is in #2941.