25956: Fixes an errant preserve_rare_values warning, hones IFA dtype validation/bugfixes, cleans up precision warnings

[apbassett]

Feature attribute suggestions

• Consolidated the three duplicated max_distilled_cases warnings in PRVSuggestion into a single _warn_default_max_distilled_cases() helper, with an optional addendum for apply().

train() dtype validation

• Numeric features are no longer flagged when held in a different numeric dtype (e.g. float64 where Int64 is expected). Previously this warned only on the nullable-integer path, while the plain int64 path silently truncated or overflowed the same data.
• Fixed tz-aware datetime columns failing validate(coerce=True): the tz check tested the Series rather than its dtype.
• Coercion failures in the categorical and datetime branches now report the underlying error instead of an empty message.
• validate(coerce=True) no longer casts a boolean column that holds nulls, which silently turned every null into False.

Excessive float precision warnings

• Features exceeding the supported 64-bit precision are collected and reported in one warning listing them all, via IFAWarningEmitter/IFAWarningCollector, instead of one warning per feature.
• IFAWarningCollector.triage now dispatches through a mapping rather than an if/elif chain.

Tests

• Added coverage for single-warning aggregation across all emitter types, the unknown-emitter-type error, and the float-precision warning.

[mkoistinen]

Approved with comments.