feat: Add Reader.RawValue for capturing a value's raw bytes

[keelerm84]

Returns the raw bytes of the next JSON value as a zero-copy slice of the
reader's input, consuming the value. Scalars are fully validated; array and
object values are scanned with a fast string-aware delimiter-matching pass.

Implemented for both the default tokenizer (position-based span capture) and
the launchdarkly_easyjson build variant (jlexer.Raw).

This enables consumers that need both a parsed representation and the
verbatim JSON of the same value (for example, the FDv2 data sources in
go-server-sdk, which store parsed items and relay raw bytes downstream) to
avoid re-scanning or copying the input.

---

Note

Medium Risk
Touches core tokenizer behavior and JSON acceptance boundaries for containers; mistakes could change what bytes are relayed downstream, though the change is additive and heavily tested against encoding/json.

Overview
Adds Reader.RawValue() and Reader.Offset() on the default build (!launchdarkly_easyjson) so callers can keep verbatim JSON alongside typed parsing without re-scanning or copying the input buffer.

RawValue consumes the next value and returns a zero-copy slice (cap == len) into the reader’s input. Scalars use the existing RFC 8259 tokenizer; arrays/objects use a string-aware delimiter scan followed by encoding/json.Valid on the captured span. Invalid containers surface a new errMsgInvalidValue syntax error. The method handles unread-token state (e.g. after a Null() probe).

Offset reports the next byte to consume (or the start of a pushed-back token), documented for the alternative pattern: slice input[start:Offset] around an in-place parse, trimming JSON whitespace at the ends.

The easyjson build intentionally omits both APIs (comment only). Coverage includes unit tests, differential/oracle tests against encoding/json, generative fuzz, and offset-span tests.

^{Reviewed by Cursor Bugbot for commit db5e127. Bugbot is set up for automated code reviews on this repo. Configure here.}