Refactor `_fixef_dict`

[leostimpfle]

The internal data structure for fixed effect coefficients and prediction doesn't take advantage of the formulaic-native encoding introduced in #1222.

Once #684 is fixed, we should refactor this to leverage vectorised prediction of fixed effects and avoid repeated loops over nested dictionaries combined with brittle string parsing.

[leostimpfle]

Here's the latest (ugly) implementation in #1222 (still based on the dictionary):

pyfixest/pyfixest/estimation/models/feols_.py

Lines 2082 to 2096 in 1235473

	fe_hat = (
	pd.concat(
	[
	fixed_effects[column].map(
	{
	float(level): coefficient
	for level, coefficient in self._fixef_dict[
	column
	].items()
	}
	)
	for column in fixed_effects.columns
	],
	axis=1,
	)