Unify `resume_from` with `initial_params` + `initial_state`

[yebai]

The functionality of resume_from is equivalent to using initial_params in conjunction with initial_state. It may be worth unifying these two mechanisms of resuming MCMC sampling, eg, by removing resume_from.

[joelkandiah]

Summary of an earlier discussion to add to this issue:
Currently:
initial_params: Resumes from an array of parameter vectors (in "constrained space")
initial_state: Resumes from an array of sampler states
resume from: resumes the model runs from an array of sampler states or from an MCMCChains object
Typically the state is resumed from chn.info.samplerstate

Ideas/Proposals:

• Remove resume_from this duplicates initial_state and constructing or extracting the states can be left to users
• Use a kwarg to state how we want the chain to be resumed and pass the information to another kwarg
  • e.g. sample(...; resume_type=:params, resume_from=init_params_array) and `sample(...; resume_type=:state, resume_from=init_states_array)'
  • This also resolves the dispatch options as users shouldn't really resume from state and params but the kwargs currently suggest that is possible. This is because typically the sampler states contain parameters.

[penelopeysm]

I definitely think it's simplest to remove resume_from, it doesn't add much except to save a few characters, imo.

[penelopeysm]

resume_from is removed in DPPL now.