fix: squeeze only axis 0 in dot_likelihood to prevent shape mismatch

[lingxiu58]

Problem

np.squeeze(X) in dot_likelihood() (pymdp/legacy/maths.py:245) removes all size-1 dimensions. When A has shape (1, N, 1) (e.g., single-modality with trailing dimension), this collapses both axes, returning (N,) instead of the expected (N, 1). This causes shape mismatches in downstream calc_free_energy() calls within fpi.py.

Fix

# Before
LL = np.squeeze(X)

# After
LL = np.squeeze(X, axis=0)

Restricts squeeze to axis 0 only, preserving the trailing dimension and maintaining shape consistency.

Reproduction

import numpy as np
from pymdp.legacy.maths import dot_likelihood

A = np.expand_dims(np.eye(25), axis=-1)  # shape (25, 25, 1)
obs = np.eye(1, 25).flatten()

LL = dot_likelihood(A, obs)
# Before: LL.shape == (25,)   ← wrong
# After:  LL.shape == (25, 1) ← correct

Fixes #171