Facilitate streaming (row-by-row update) in ESMDA

[tommyod]

To support row-by-row (parameter by parameter) computation, either literally by updating a single parameter or updating batches, we should avoid doing duplicate work.
Consider the equation:

$$C_{MD} (C_{DD} + \alpha C_D)^{-1} (D - Y)=X Y^T /(N-1) (C_{DD} + \alpha C_D)^{-1} (D - Y)$$

If we update row by row and introduce a transition matrix $K := Y^T /(N-1) (C_{DD} + \alpha C_D)^{-1} (D - Y) $,
then $[x_1 | x_2 | \ldots]^T K = [x_1 K | x_2 K | \ldots ]^T$. In other words, we can pre-compute $K$, since it's independent on $X$, and apply it row-by-row or group-by-group, instead of re-computing it for each group.

---

This PR proposes two methods for this:

• compute_transition_matrix, which computes the transition matrix $K$
• perturb_observations, called by both compute_transition_matrix and assimilate, to avoid duplicating code

Note: this PR assumes #161 will be merged.

---

I also realized that we don't have to center both $X$ and $Y$! So I changed the code away from that, saving both memory and time by only centering the smaller matrix $Y$. See this comment. This is also documented in the code with a few lines.

[tommyod]

Remove this

Suggested change

	X - np.mean(X, axis=1, keepdims=True)
	X - np.mean(X, axis=1, keepdims=True)

[Blunde1]

LGTM.
This is very nice! In particular, see comparison between low-level and high-level API in tests. Good job!
Note for later: the names inversion are slightly missleading, they all involve inversion, but their purpose is either to return K or multiply X with K. Perhaps we can come up with good naming. As discussed, they are not part of the public api, so we may think of it later.