[numba.md] numba_ex3: release the 1e8 arrays after the timings

The `u_big`/`v_big` pair added for numba_ex3 stayed alive for the rest of
the notebook. Because `numba_ex_draw_speed` rebinds `u_draws`/`v_draws` to
a second pair of 1e8-element arrays, peak memory reached ~3.2 GB rather
than ~1.6 GB. That is comfortable on the g4dn.2xlarge CI runner but not on
the smaller machines the accompanying memory note is written for.

Delete the arrays after their last use, and fix "Lets" -> "Let's" plus the
trailing whitespace introduced alongside it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: mmcky

lectures/numba.md

@@ -723,10 +723,10 @@ def calculate_pi_parallel(u_draws, v_draws):
     return area_estimate * 4  # dividing by radius**2
 ```
 
-Parallelization pays off when each thread has enough work to overcome the overhead costs while 
-breaking the problem into parts for work to happen simultaneously. 
+Parallelization pays off when each thread has enough work to overcome the overhead costs while
+breaking the problem into parts for work to happen simultaneously.
 
-Lets draw a fresh, much larger set of points rather than reusing the arrays from above.
+Let's draw a fresh, much larger set of points rather than reusing the arrays from above.
 
 ```{note}
 The two arrays below occupy about 1.6 GB of memory — reduce `n` if your
@@ -769,6 +769,13 @@ on the number of CPUs on your machine — and at small sample sizes the
 parallel version can even be slower, because the gains cannot cover the cost of
 distributing work across threads.)
 
+These two arrays are large and we are finished with them, so we release the
+memory before moving on.
+
+```{code-cell} ipython3
+del u_big, v_big
+```
+
 Notice that we drew all of the random points *before* the loop and passed them in
 as arrays, so the parallel loop only *reads* from memory.