Fix sign error in Q-learning gradient-descent update equations

[Chessing234]

Closes #2649.

Bug

In chapter_reinforcement-learning/qlearning.md (section 17.3), the expanded form of the gradient-descent Q-learning update has a flipped sign:

Q <- Q - alpha * grad_Q l(Q)
    = (1 - alpha) Q - alpha ( r + gamma max_a' Q' )    <-- wrong

The same flipped sign also appears in the next equation (terminal-state variant).

Root cause

The loss is l(Q) = (Q - r - gamma max_a' Q')^2, so grad_Q l = 2 (Q - r - gamma max_a' Q'). Substituting into the gradient step and absorbing the factor of 2 into alpha:

Q - alpha * (Q - r - gamma max_a' Q')
  = Q - alpha*Q + alpha*(r + gamma max_a' Q')
  = (1 - alpha) Q + alpha (r + gamma max_a' Q')

The book's expansion kept the minus sign from the gradient-descent step instead of distributing it through the parentheses, producing (1 - alpha) Q - alpha (...).

Why the fix is correct

The in-chapter Python implementation at line 137 of the same file already uses the correct form:

y = reward + gamma * np.max(Q[next_state,:])
Q[state, action] = Q[state, action] + alpha * (y - Q[state, action])

Expanding: (1 - alpha) Q + alpha * y, matching the fixed equation. The text and the code were inconsistent; this commit makes the text match the code (and the standard Q-learning update rule).

Change

Two single-character edits (- → +) on the two equation lines. No other changes.