styling: validate init_sd / column_config_overrides at the entry points (fast follow of #978)

[paddymul]

Problem

Fast follow to #978. Once the display-config keys have a type, the same keys need checking at runtime, because most of them arrive from places no type checker ever sees: notebook calls, JSON on the wire, and project display files loaded through exec (server/xorq_loading.py).

Today a malformed value either raises inside style_column — costing that column its styling — or is silently ignored. After #977 the two known cases are ignored: a non-dict ag_grid_specs and a non-list delete_keys (customizations/styling.py:137,143). Silently ignoring a typo conflicts with wanting a broken setup to be obvious; it just moves the failure from loud-and-wrong to quiet-and-wrong.

Suggested fix

Check init_sd and column_config_overrides where they enter, rather than at the point of use:

• each config key holds the shape its type says (displayer_args / ag_grid_specs dicts, delete_keys a list of strings, highlight_* strings).
• report what was wrong and which column it was on, rather than dropping it. A warning naming the column and the key is the minimum; a channel that reaches the browser is better, and shares its mechanism with styling: surface style_column exceptions in the column config instead of only in server logs #971.
• a bare delete_keys: 'tooltip_config' is accepted as the single key (fix(styling): make base styling survive malformed init_sd, keep default_styling as the fallback hook #977 already does this) — it's a typo with an unambiguous intent, unlike the others.

Keep the guards at the point of use regardless: entry validation can't cover a ColAnalysis subclass that contributes a bad value into the sd from inside the pipeline.

Context

Split out of #978 as the runtime half. Related: #976 (validating the configs styling produces, i.e. the other end of the same pipe), #977 (the guards that prompted this), #971.