feat(instruction_following): add loose-mode evaluation and compute_metrics

[rodboev]

The instruction_following server currently evaluates each instruction in strict mode only and returns no aggregate metrics from compute_metrics(). As a result, rollouts_aggregate_metrics.json omits all four IFEval accuracy dimensions: prompt_strict_accuracy, instruction_strict_accuracy, prompt_loose_accuracy, and instruction_loose_accuracy.

This PR adds loose-mode evaluation and the missing compute_metrics() override:

• Adds follow_all_instructions_loose and follow_instruction_list_loose to InstructionFollowingVerifyResponse
• Adds a _check_following_loose helper that uses a native loose API if available (check_following_loose / check_following(..., mode="loose")) and falls back to the IFEval perturbation set: the original text, without-first-line, without-last-line, and without-first-and-last-line variants, each duplicated with asterisks removed, matching the NeMo Skills convention
• Computes both strict and loose from the same instruction instance in a single loop inside verify(), avoiding double instantiation
• Implements compute_metrics(tasks: List[List[Dict]]) on InstructionFollowingResourcesServer, matching the base class signature, and aggregates all four IFEval accuracy scores as percentages (0-100), consistent with compute_pass_majority_metrics
• Adds focused tests covering loose field presence, the loose >= strict invariant, and compute_metrics() arithmetic

No new dependencies. verifiable_instructions is already in requirements.txt.

Closes #1160

Checklist

• Targeted tests pass: uv run pytest resources_servers/instruction_following/tests/test_app.py -x -v
• Touched-file lint passes: ruff check --config pyproject.toml resources_servers/instruction_following/app.py resources_servers/instruction_following/tests/test_app.py
• Touched-file format check passes: ruff format --config pyproject.toml --check resources_servers/instruction_following/app.py resources_servers/instruction_following/tests/test_app.py
• No changes to benchmark configs, data files, or nemo_gym core
• DCO sign-off on all commits

Targeted tests were not run locally; upstream CI is the source of truth.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.