[Perf][C++] Avoid memcpy of the RLE pattern in read_length_and_bitwidth #1076

Problem

Int64RleDecoder::read_length_and_bitwidth / Int32RleDecoder::read_length_and_bitwidth
copy the whole encoded pattern into a temporary buffer (tmp_buf_) on every
pattern, then wrap byte_cache_ over that copy. Profiling a wide-table read
showed this memcpy at ~7.5% of the query.

Proposed changes

When the caller's ByteStream is wrapped and has at least length_ bytes left,
wrap byte_cache_ directly over the caller's buffer and advance its read
position; fall back to the copy for paged or short streams. The caller's stream
outlives the pattern decode, so no lifetime issue.

Implemented in #1072.

Related: #1072