JSON wire format for Expr values.

The Python bindings build expression trees as JSON; this module decodes them back into Expr GADTs (typed) and encodes them in the reverse direction (used for shipping fitted decision trees back to Python).

Encoded shape:

{ "node": "col"   , "out_type": "double", "name": "x" }
{ "node": "lit"   , "out_type": "int"   , "value": 42 }
{ "node": "unary" , "out_type": "double", "op": "toDouble"
                  , "arg_type": "int"   , "arg": <expr> }
{ "node": "binary", "out_type": "bool"  , "op": "leq"
                  , "arg_type": "double", "lhs": <expr>, "rhs": <expr> }
{ "node": "if"    , "out_type": "text"  , "cond": <expr>
                  , "then": <expr>, "else": <expr> }

Operator names follow binaryName / unaryName from DataFrame.Internal.Expression. Nullable variants ("nulladd", "nullor", …) are accepted on decode and treated as their non-null equivalents — Python data crosses the FFI boundary as non-null Arrow buffers, so this is lossless for the supported workflow.