[RFC 0005] Reconstruct `BARTSamples` / `BCFSamples` from JSON

[andrewherren]

• Envelope write (R/Python): gather scalars/params from the held BART/BCFSamples, serialize forests via the existing C++ ForestContainer serialization, serialize random effects similarly, assemble the unified envelope, and embed the language's native covariate_preprocessor + portability header (the native preprocessor stays serialized by the language that owns it).
• Envelope read (R/Python): parse the unified envelope; load forests via the existing procedures; rebuild the preprocessor (if cross-platform be sure to handle all-numeric instructions correctly).
• BART/BCFSamples construction: populate/construct an in-memory samples object from R/Python-provided forest containers + parameter arrays (copying parameter array, moving pointers).
• Sampler takes BARTSampler(BARTSamples&, …) / BCFSampler(BCFSamples&, …) -- continued sampling operates on the in-memory object; from-disk = read -> populate -> resume.
• Round-trip parity: write -> read -> identical in-memory state (both languages).
• Cross-platform: an R-written numeric model populates a BARTSamples in Python (and vice versa) and predicts / continues.