openmetrics: escape sample metric names consistently with HELP/TYPE

[amdadulbari]

Summary

Fixes #1177.

In the OpenMetrics exposition, the # HELP / # TYPE / # UNIT lines escape the metric name with escape_metric_name() (where : is a valid metric-name character and is kept), but the sample line escaped the name with the label-name rune set (_is_legacy_labelname_rune), which turns : into _.

So a metric named sglang:token_usage was emitted as:

# HELP sglang:token_usage Total token usage.
# TYPE sglang:token_usage gauge
sglang_token_usage 42.0

The metadata name and the sample name disagree, which strict OpenMetrics parsers (e.g. promtool check metrics) reject as an orphaned metadata set plus an untyped metric.

Change

Use escape_metric_name() for the sample name as well, so a single helper drives HELP, TYPE, UNIT and the sample line for every escaping scheme (underscores, dots, values, allowutf8). This matches the existing escape_metric_name semantics already asserted by test_escape_metric_name (a colon is preserved in metric names).

Tests

Added test_metric_name_with_colon_consistent_across_metadata_and_samples, which checks the exact underscores-escaped output and asserts the metric-name identifier is identical across HELP, TYPE and the sample line for every escaping scheme. It fails on master and passes with this change. Full tests/openmetrics, tests/test_exposition.py and tests/test_parser.py suites pass; flake8 and isort are clean.