fix: reject nulls at load time in every schema validator

[michael-denyer]

Closes the one follow-up left open by #25.

validate_gwas_dataframe already required non-null positions and p-values.
validate_eqtl_dataframe, validate_finemapping_dataframe, and
validate_genes_dataframe did not, so a file with a blank p_value, pip,
start, or end passed load validation and silently lost those rows later,
when plot-time intake filtered them with a warning the user was unlikely to
connect back to the file.

That gap was an artefact of the old hand-rolled implementation rather than a
decision. #25 preserved it deliberately so the refactor stayed
behaviour-neutral; this is the labelled behaviour change that closes it.

Change

All four schema validators now reject nulls in the numeric columns they already
range-check, naming the offending column:

Validator	Columns newly checked
validate_eqtl_dataframe	pos, p_value
validate_finemapping_dataframe	pos, pip
validate_genes_dataframe	start, end

validate_gwas_dataframe is unchanged; it was already strict.

What is deliberately not changed

Plot-time intake stays lenient and still filters nulls with a warning
(prepare_pvalue_data, prepare_eqtl_for_plotting). That split is the point: a
frame reaching the plotter may have been assembled by the caller rather than
read from a file, so the strict gate belongs at the file boundary. Recorded in
CONTEXT.md next to the other load-bearing intake distinctions, and in the
contract test's module docstring.

Verification

Five new contract cases were red before the change and green after, and two
cases that asserted the old tolerance are flipped to rejections.

End to end through a real loader, not just the validator:

clean SuSiE file: 2 rows loaded
null pip file: rejected -> Fine-mapping validation failed:
  - Column 'pip' has 1 null values

Full suite: 1195 passed, 1 skipped. ruff check and ruff format clean.

Release note

2.0.0 is merged but not yet tagged, so this lands in the existing 2.0.0
CHANGELOG entry rather than a patch bump.