Use StringLookaheadBuffer if a string is available.

[maxvoxel8]

Performance improvement to use StringLookaheadBuffer where a string can be directly accessed. This proves important especially with YAML that has a lot of quoted values that makes heavy use of ScanFlowScalar.

This is a before and after comparison in DotTrace using Sample mode:

This is using Tracing mode:

[maxvoxel8]

Also don't allocate the extra StringBuilders in ScanPlainScalar unless needed.

[maxvoxel8]

• Speed up enumerating ParsingEvents of YamlNode hierarchy by creating a custom enumerator (don't have profiler screenshot but it sped up my test case by > 1 second).
• Speed up parsing by replacing mutable Mark struct with inline variables.

[maxvoxel8]

The left profile is before my changes, the middle profile is with 6f5e365, the right is with 5a168a5.

[maxvoxel8]

Speed up adding nodes when tracker is used. This is a performance penalty that only exists if tracker is assigned.

[maxvoxel8]

Improve write performance by reducing allocation in AnalyzeScalar (ideally the StringLookaheadBuffer should also not be allocated, but that requires more changes).

[maxvoxel8]

Speed up YamlNode.DeepClone by not copying immutable ParserEvent values and by pre-allocating list size. Not sure why I was copying the events to begin with...

[maxvoxel8]

@xoofx thanks for the merge, I do have another performance change I'm working on that I will make another PR for.