fix(gene burden): normalise NaN effect sizes to null

[vivienho]

Summary

• Normal rows leave beta/oddsRatio as SQL NULL when no effect size is reported, e.g. most gene_burden rows correctly get directionOnTrait: NULL when both fields are null.
• SKAT/SKAT-O rows from BRaVa and Genes & Health instead got directionOnTrait: 'risk' even though no effect size was ever computed, e.g. beta: NaN, oddsRatio: null.
• The source files encode "no effect size" as the string "NaN" ('BETA Burden' in BRaVa's S14/S15, 'effect'/BETA in Genes & Health's ST9/ST13/ST15), which Spark's double cast turns into a float NaN rather than NULL.
• Spark SQL orders NaN as larger than any other value, so direction_on_trait_expression's beta > 0.0 (or oddsRatio > 1.0) evaluates TRUE for a NaN input, misclassifying 5,617 rows as 'risk'.

Fix

• In _process_brava_granular, normalise 'BETA Burden' from NaN to NULL right after casting it to double, so SKAT/SKAT-O rows (which carry no effect size) behave as missing.
• In process_genes_and_health_gene_burden, normalise 'effect' from NaN to NULL right after unioning the three source tables, before it's used to derive beta/oddsRatio.
• Corrected output: beta: null, oddsRatio: null → directionOnTrait: null instead of 'risk'.

[ireneisdoomed]

Well spotted! I have double checked that the columns with NaN in the numeric columns are: beta, oddsRatio + the correspondent 4 fields about the upper and lower bounds of the confidence intervals.

Because you are standardising the effect and beta columns directly, the derived columns will be fixed too.