Scaffold `glm` module

[sdiebolt]

Describe the new feature you'd like

Create the confusius.glm module for general linear model analysis of fUSI data,
following the nilearn design pattern.

Planned scope (alpha)

• FirstLevelModel: fit a voxel-wise GLM given a design matrix (stimuli
  convolved with an HRF, confound regressors) and return contrast maps as
  xr.DataArray.
• SecondLevelModel: fit a group-level GLM over a collection of first-level
  contrast maps (e.g., fixed-effects or mixed-effects across subjects/sessions).
• Design matrix helpers: build first-level design matrices from stimulus
  paradigms (onset/duration/amplitude), HRF convolution, and confound columns.