feat(predict-rlm): add in-context string inputs

[magix022]

Rationale

Some PredictRLM tasks need small-to-moderate string inputs, such as rubrics or criteria, to be visible directly to the outer RLM instead of only being discoverable through REPL variable inspection. This adds a typed way to opt into that behavior while preserving the existing variable-based workflow.

Summary

• Add InContext, a public string input marker exported from predict_rlm.
• Append marked input values verbatim to the action and extract prompts under an In-Context Inputs block.
• Keep the same values available as normal Python string variables in the REPL.
• Document the new input type and add focused tests for prompt formatting, validation, and signature composition.

Test Plan

• uv run pytest tests/test_in_context.py tests/test_shared.py::TestBuildRlmSignatures::test_in_context_instructions_appended_last_to_action_and_extract -q
• uv run ruff check src/ tests/
• make test-unit