Add ConvergentFilter

[jessegrabowski]

Closes #394
Related to #332

For certain classes of models, the latent state covariance matrix estimated by the Kalman Filter converges to a fixed point, after which is is no longer necessary to compute via expensive linear algebra. This PR adds a new filter, ConvergentFIlter, that exploits this property. To take advantage, a model must:

• Have no missing values, and
• No time varying matrices

The ConvergentFilter uses a while scan, so this cannot be used with jax. All timings are done with numba.

These break the convergence assumptions. But nothing else is required. If these criteria are met, we get big speedups. Here is a table of models, where n is the number of timesteps, m is the number of hidden states, p is the number of observed states, and r is the number of shocks. The biggest speedups come from long time series, where there is more iterations in the compute-saving post-convergence scan. Here are some timings for the compiled logp_dlogp function for models taken from the example notebooks:

Model	Shape (m, p, r)	n	StandardFilter	ConvergentFilter	Speedup
VARMAX(2,0) 5-obs	(10, 5, 5)	200	3.4	1.3	2.5x
VARMAX(2,0) 5-obs	(10, 5, 5)	1000	17.2	5.1	3.4x
VARMAX(2,0) 5-obs	(10, 5, 5)	5000	96.0	26.6	3.6x
SARIMAX(1,0,1)	(2, 1, 1)	200	1.1	0.6	1.9x
SARIMAX(1,0,1)	(2, 1, 1)	1000	5.4	2.1	2.6x
SARIMAX(1,0,1)	(2, 1, 1)	5000	27.9	10.1	2.8x
SARIMAX(2,1,2)(2,0,2,12)	(28, 1, 1)	200	5.7	3.1	1.8x
SARIMAX(2,1,2)(2,0,2,12)	(28, 1, 1)	1000	30.7	14.8	2.1x
SARIMAX(2,1,2)(2,0,2,12)	(28, 1, 1)	5000	153.7	63.9	2.4x
ETS(A,Ad,A,12)	(15, 1, 1)	200	3.4	1.8	1.9x
ETS(A,Ad,A,12)	(15, 1, 1)	1000	17.6	6.4	2.8x
ETS(A,Ad,A,12)	(15, 1, 1)	5000	90.7	33.6	2.7x
DFM 1-factor 2-lag 4-obs	(10, 4, 5)	200	2.9	1.5	1.9x
DFM 1-factor 2-lag 4-obs	(10, 4, 5)	1000	15.6	4.9	3.2x
DFM 1-factor 2-lag 4-obs	(10, 4, 5)	5000	78.3	24.8	3.2x
Structural LvlTrend+FreqSeas (airpass)	(14, 1, 14)	200	2.5	1.5	1.7x
Structural LvlTrend+FreqSeas (airpass)	(14, 1, 14)	1000	12.8	4.6	2.8x
Structural LvlTrend+FreqSeas (airpass)	(14, 1, 14)	5000	66.5	23.3	2.9x

Special thanks to @JeanVanDyk, who derived components of the closed forms used in the handoff step between the two scans. I tried using the full analytic forms he worked out end-to-end, but it was slower than autodiff, I believe because of the number of matmuls involved. It might be faster on GPU, so I want to revist that again at some point.