fix(studio): cross-model comparison view in Studio UI

[christso]

Objective

Enhance the Studio UI to provide side-by-side comparison of eval results across different targets/models within the same experiment.

Motivation

The agentv CLI has a compare command, but the Studio UI currently shows runs individually. When running the same eval across different providers (e.g., azure vs gemini) with the experiment feature (#977), users need to compare results side-by-side.

From benchmarking work: running reasoning evals across azure (gpt-5.4-mini) and gemini (gemini-3-flash-preview) with with-superpowers/without-superpowers experiments produces 4 separate runs. The Studio should enable:

• Seeing all 4 runs in a comparison matrix
• Identifying which target + experiment combination performs best
• Spotting grading anomalies (e.g., "no response provided" grading failures)

Design

Extend the existing experiments tab in Studio to show a comparison matrix:

                    | without-superpowers | with-superpowers |
azure (gpt-5.4)    | 73.5% (1/2 pass)   | 25.0% (0/2 pass) |
gemini (flash)     | 100% (2/2 pass)    | 75.0% (1/2 pass)  |

Implementation approach

• Add a "Compare" view to the experiments tab
• Group runs by experiment × target
• Show pass rate, average score, and per-test-case breakdown
• Highlight best/worst performers
• Support drill-down to individual test case differences

Acceptance Criteria

• Experiments tab shows comparison matrix when multiple targets exist
• Pass rate and average score visible for each experiment × target cell
• Drill-down to per-test-case comparison
• Works with existing experiment-based results layout

Non-goals

• Statistical significance testing (future enhancement)
• Automated recommendations