Add VI for linear regression

[ShouvikGhosh2048]

No description provided.

[ShouvikGhosh2048]

This seems like a very good idea.

@sourish-cmi @Nandini-Jaiswal

I've added VI for linear regression.

• I split BayesianRegression into BayesianRegressionMCMC and BayesianRegressionVI structs.
• I pass a boolean use_vi to determine whether to use MCMC or VI.
• For VI, I store the distribution, and sample it during prediction. (This technically makes the prediction random each time, I could change it to sample during creation and store the sampled values instead dist, if you prefer that.)
• I made seperate structs for BayesianRegressionMCMC and BayesianRegressionVI to store the chain and dist respectively. However we could just convert both to a matrix of samples (we will sample the dist for VI) and just have a single BayesianRegression struct which stores the matrix of samples. We would lose the info about the chain and dist though, I'm not sure if its important.

[sourish-cmi]

Looks like very good improvement.

[sourish-cmi]

@ShouvikGhosh2048 Can you please add tests and check why the doc test is failing? Do you want me to provide some test!

[ShouvikGhosh2048]

I've added tests for VI, those are passing (right now just whether they run, I'll add a test to make sure that the MCMC and VI predictions are close). The doc tests are failing since I haven't made the changes for VI, I'll make them once we decide on the API.

[sourish-cmi]

Okay - can we meet over Google Meet?

[ShouvikGhosh2048]

I've made the changes we discussed and updated the docs.

I ended up removing use_vi and instead I use a BayesianAlgorithm struct defined here:
https://github.com/xKDR/CRRao.jl/pull/154/files#diff-b989ddbe26d5a6bf4e0714d4bc169310463a8e8bbcadaf6d5903d5fe34f6fc02R395-R432

I went with this since the algorithms have different parameters (MCMC needs the number of iterations and the number of initial observations we discard, and VI needs the number of samples, number of VI iterations and number of samples per VI step).

So now we do:
model = fit(@formula(MPG ~ HP + WT + Gear), mtcars, LinearRegression(), prior) - MCMC
model = fit(@formula(MPG ~ HP + WT + Gear), mtcars, LinearRegression(), prior, VI()) - VI

If the API is fine, this PR is ready to merge.