feat: fast EntropyBottleneck aux_loss minimization via bisection search

[YodaEmbedding]

This method completes in <1 second and reduces aux_loss to <0.01.

This makes the aux_loss optimization during training unnecessary.

Another alternative would be to run the following post-training:

while aux_loss > 0.1:
    aux_loss = model.aux_loss()
    aux_loss.backward()
    aux_optimizer.step()
    aux_optimizer.zero_grad()

...but since we do not manage aux_loss learning rates, the bisection search method might converge better.

Note: I haven't yet tested this extensively, so there might be some rough edges.

[YodaEmbedding]

Why are the automated tests failing?

Previously, update() did not modify the default quantiles, which are initialized to [-10, 0, 10]. With this PR's modification to update(), it automatically updates the quantiles, too. But then, the initial quantiles become e.g. [-216, -0.4, 207], so that median=-0.4, and thus y = -0.6 is quantized to y_hat = -0.4. However, the tests were written with the assumption that median=0, and thus they expect y_hat = -1.

    def test_compression_2D(self):
        x = torch.rand(1, 128, 32, 32)
        eb = EntropyBottleneck(128)
        eb.update()
        s = eb.compress(x)
        x2 = eb.decompress(s, x.size()[2:])

        # THE torch.round(x) ASSERT ASSUMES THAT:
        # assert (eb.quantiles[..., 1] == 0).all()

        assert torch.allclose(torch.round(x), x2)

Possible fixes:

1. Patch the tests to disable auto-updating the quantiles via eb.update(update_quantiles=False). (The "easy" way.)
2. Fix the tests to use torch.round(x - means) + means instead. (The "correct" way.)

I went with option 2.