Define unique: NaN handling in table schema validation (treat NaN as equal)

[nick-gorman]

Context

Several table schemas declare uniqueness over a key that includes a nullable
column (e.g. date_from, timeslice). Standard SQL UNIQUE treats NULLs as
distinct, so a plain unique: over such a key would not enforce "at most one
NaN row per group". To get that behaviour we currently use bespoke
custom_validation rules:

• custom_constraints_lhs.yaml -> no_duplicate_constraint_term_date
• custom_constraints_rhs.yaml -> no_duplicate_constraint_timeslice_date
• network_transmission_path_limits.yaml -> no_duplicate_path_direction_timeslice

In each case NaN is a meaningful sentinel (e.g. date_from absent = "applies
from the start of the horizon"), and we want at most one such row per group.

Working decision

unique: treats NaN as equal to NaN (pandas duplicated() semantics): rows
that are equal on the key apart from NaN values are still duplicates, so at most
one otherwise-identical NaN row is permitted. This makes a plain unique:
declaration enforce what the custom rules above describe.

Open to revisiting if anyone objects.

Follow-up

• Replace the three custom_validation rules above with plain unique:
  declarations now.
• Decide where to document the unique: NaN contract (schema-format reference
  and/or validator docstring) if we keep this decision.