Implementation of fast likelihood with MarginalizingTimingModel

[vallis]

This PR implements the same algebra as #286, but confines it to a smart MarginalizingTimingModel object that does two things:

• Collects all other white-noise objects that would normally make up the N matrix, and stores their sum internally.
• When queried with get_ndiag(), returns a MarginalizingNmat object with a solve() method that returns (r^T D^-1 r), (T D^-1 r), (T D^-1 T^t) for the matrix D = N + M E M^T (with E formally infinite variance).

All operations are cached appropriately. This means that the TimingModel will not participate in the GP basis and Phi inversion. Note that MarginalizingTimingModel should be placed first when assembling a model.

The PR also modifies the standard Loglikelihood to accept a parameter cholesky_sparse=False that disables the sparse Cholesky of the combined Sigma matrix, and that caches the combined TNr and TNT. These steps are required to get the full benefits of MarginalizingTimingModel.

[codecov[bot]]

Codecov Report

Merging #288 (7e0192b) into master (a06d1f6) will increase coverage by 0.32%. The diff coverage is 96.90%.

@@            Coverage Diff             @@
##           master     #288      +/-   ##
==========================================
+ Coverage   86.44%   86.76%   +0.32%     
==========================================
  Files          12       12              
  Lines        2730     2819      +89     
==========================================
+ Hits         2360     2446      +86     
- Misses        370      373       +3     

Impacted Files	Coverage Δ
enterprise/signals/gp_signals.py	88.00% <95.38%> (+1.24%)	:arrow_up:
enterprise/signals/signal_base.py	89.75% <100.00%> (+0.33%)	:arrow_up:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update a06d1f6...7e0192b. Read the comment docs.