Fix Pi reasoning effort options to respect model capabilities

[lcisconl]

Fixes #395.

Pi previously offered a fixed set of reasoning efforts, including levels the selected model does not support. Build the options from Pi's thinkingLevelMap: hide levels mapped to null, preserve ordinary defaults for omitted keys, and require explicit support for xhigh and max.

The existing Pi model adapter remains the owner of this behavior. The shared pickers consume the filtered model options, including when they are supplied by a remote host. There are no model-name special cases, new dependencies, or protocol changes. The composer design contract is updated.

Regression tests cover absent and partial mappings, disabled levels, extended levels, unsupported defaults, and non-reasoning models. The filtering regression fails with the original map_model function and passes with the fix. A separate non-reasoning test replaces repeated assertions for the same contract.

Validation on Windows with Rust 1.98.1:

• cargo fmt --all --check: passed.
• cargo build --workspace --locked: passed.
• cargo clippy --workspace --all-targets --locked -- -D warnings: passed.
• cargo machete 0.9.2: passed.
• Pi tests: 15 passed; the workspace run also passed all 174 agent tests.
• cargo test --workspace --locked: stopped at two runtime failures, resident_idle_reaper_shuts_down_untouched_provider and orchestrate_worktree_dispatch_resolves_child_cwd_to_worktree (temporary worktree deletion returned Permission denied). The full suite is not green.
• Individual reruns: the idle-reaper test passed; the worktree-deletion failure reproduced.

Checks were run on the patch over cba07d3. macOS, Linux, mobile/Web target checks, and two-device acceptance were not run.