feat(metrics): add metric validation + UTF-8 support

[ethanolchik]

What Changed

Introduces shared validation used during collection and by both encoders.

Names are valid when they are non-empty UTF-8 strings without NUL (\0) bytes. Names containing spaces, punctuation, newlines, or non-ASCII characters remain valid and are escaped appropriately for text output.

Validation failures are isolated as narrowly as possible:

• An invalid family name drops that family.
• An invalid, duplicate, or reserved label drops only its metric row.
• An invalid or duplicate exemplar label drops only the exemplar.
  • This branch validates exemplars already present in MetricFamily; exemplar-producing metric types are introduced separately.
• Other families and rows continue to be collected and encoded.
• Every dropped value produces a non-fatal diagnostic with collection or encoding context.

Histogram and gauge-histogram rows reserve le, while summary rows reserve quantile, according to the canonical metric model.

Service Labels

Service-label insertion now handles existing labels safely:

• A row already containing the configured label and value is unchanged.
• A row containing the same label with a different value is dropped.
• A missing service label is prepended as before.
• An invalid configured service-label name prevents collection instead of producing malformed output.

OpenMetrics Text

The text encoder now:

• Quotes label names only whne they require UTF-8 syntax; legacy-compatible label names remain unquoted.
• Uses quoted OpenMetrics syntax for metric names outside the legacy grammar.
• Escapes metric and label names safely.
• Appends histogram suffixes inside quoted metric names.
• Exports OPENMETRICS_CONTENT_TYPE with escaping=allow-utf-8.
  Legacy-compatible metric names remain unquoted.

Protobuf

Protobuf output preserves valid UTF-8 names exactly. The protobuf encoder applies the same defensive validation when called directly, so bypassing collection does not bypass validation.