Robust BART from_json: safe defaults and warnings for missing fields

[andrewherren]

Part of #315 (cross-version deserialization, RFC 0002).

Update BART from_json() (Python) and the R BART deserialization functions to check for each optional field before reading it, falling back to a safe default and emitting a warning that names the field, the assumed default, and the inferred version bracket (from #318).

Missing field	Safe default
covariate_preprocessor / preprocessor_metadata	None / NULL — warn; disables DataFrame prediction
rfx_model_spec	"" / None — warn if has_rfx = TRUE
has_rfx_basis	FALSE / False
num_rfx_basis	1
outcome (in outcome_model)	"continuous"
link (in outcome_model)	"identity"
probit_outcome_model	FALSE / False
cloglog_num_categories	0
num_chains	1
keep_every	1

Hard errors should be reserved for genuinely unrecoverable fields (forest structures, outcome_scale, outcome_mean).

Affects: stochtree/bart.py, R/bart.R