[experiments] Metric charts on the experiment compare page

[mikeldking]

Part of #11637.

Add metric charts to the experiment compare page's metrics view (?view=metrics, ExperimentCompareMetricsPage), styled like the charts on the experiments page:

• Per-evaluator charts: one chart per evaluator present on the selected experiments. Numeric evaluators show the mean score; label evaluators show the label distribution
• Latency: average run latency per experiment
• Errors: share of runs that errored per experiment
• Token breakdown: prompt tokens and completion tokens per experiment, stacked by token type

Context

The experiments page charts (pages/dataset/metrics/chartCatalog.tsx) get their data from useExperimentMetricsData(datasetId). That hook always loads the dataset's last EXPERIMENT_METRICS_EXPERIMENT_COUNT experiments plus the dataset baseline. On the compare page, the charts need to show only the selected experiments and use the base experiment as the reference.

Scope

Data

• Load chart data for the selected experiment IDs (base + compare). Dataset.experiments(filterIds:, includeEphemeral: true) already supports this; the compare metrics query uses it.
• Let the chart views accept their experiments + baseline data from either source (dataset recent experiments or compare selection) without forking each chart.
• Get evaluator names from the selected experiments, not from the whole dataset.
• Charts update when compared experiments are added or removed, without a full reload.

Presentation

• On the x-axis, order experiments base first, then compare experiments in selection order. Color each one with useExperimentColors so the charts match the rest of the compare page.
• Draw the base experiment as the reference, using the same treatment as the dataset baseline (ExperimentBaselineReference).
• Latency: mirror ExperimentLatencyChart.
• Errors: mirror ExperimentErrorRateChart. Hovering shows error rate and run count.
• Token breakdown: mirror ExperimentPromptTokenDetailsChart / ExperimentCompletionTokenDetailsChart, using data from costDetailSummaryEntries. Use the labels and colors from tokenDetailUtils, and key details off the raw token type.
• Evaluators: reuse ExperimentAnnotationMetricPanel. Defer offscreen charts (DeferredChartPanel) so many evaluators don't all fetch up front.

Acceptance

• Metrics view shows latency, error rate, prompt token, and completion token charts for exactly the selected experiments
• One chart per evaluator present on the selected experiments; numeric and label evaluators each render their appropriate view
• An experiment missing an evaluator or metric shows as no data, not as zero
• The base experiment is visually marked as the reference
• Experiment colors match the compare page's experiment colors
• The experiments page charts behave as before
• Storybook story covers the compare chart variant