Fix surrogate pairs mis-decoded when split across a read chunk boundary - #4478

Parser.setupCharStream() used ANTLR's CharStreams.fromReader(), which reads
the document in fixed 4096-char chunks. CodePointBuffer.Builder keeps a
pending high surrogate across chunks to support pairs split by a chunk
boundary, but its end-of-chunk flush fails to clear that pending state,
so the surrogate gets decoded twice: once as a lone code point at the
chunk boundary, and again when paired with the next chunk's low surrogate.

In practice this makes graphql-java reject an otherwise valid document
whenever an astral character (e.g. an emoji) lands with its high surrogate
at index 4095, 8191, etc., since the resulting lone surrogate fails
GraphQL's StringCharacter grammar rule.

This is a known upstream defect (antlr/antlr4#4943, open and unreviewed).
Rather than wait on it, this replaces fromReader() with an equivalent local
read loop that holds back a chunk-trailing unpaired high surrogate so it is
carried into the next chunk instead of being flushed prematurely.