Change regex usage to manual parsing

[lahma]

I started profiling the "worst case" of evaluating formulas which seemed to take a lot of resources. The culprit turned out to be usage of regex parsing to get column/row/cell information. I created a custom parser that should fulfill the current needs and is a lot faster as it's being called in hot code paths. I also updated APIs to allow also ReadOnlySpan<char> for operations that needed substring to work with.

Based on running the benchmarks in the other PR, here are the improvement numbers. Still room to improve but this was the low-hanging fruit as first step.

NPOI.Benchmarks.LargeExcelFileBenchmark

Diff	Method	Mean	Error	Allocated
Old	Load	11.741 s	1.1257 s	3.65 GB
New		11.497 s (-2%)	2.4215 s	2.79 GB (-24%)
Old	Write	3.483 s	0.1738 s	1.19 GB
New		3.492 s (0%)	0.2931 s	1.19 GB (0%)
Old	Evaluate	62.677 s	4.1891 s	75.32 GB
New		18.854 s (-70%)	2.7709 s	12.93 GB (-83%)

NPOI.Benchmarks.RangeValuesBenchmark

Diff	Method	Mean	Error	Allocated
Old	Double	85.46 ms	0.367 ms	31.45 MB
New		86.43 ms (+1%)	0.181 ms	31.45 MB (0%)
Old	String	99.93 ms	0.449 ms	59.32 MB
New		98.76 ms (-1%)	0.521 ms	59.32 MB (0%)
Old	Date	114.71 ms	1.603 ms	35.73 MB
New		115.39 ms (+1%)	0.832 ms	35.73 MB (0%)
Old	Formulas	1,064.12 ms	10.889 ms	1480.15 MB
New		402.57 ms (-62%)	1.058 ms	284.07 MB (-81%)

[tonyqus]

Wow, this is a incredible result. I'm looking forward to this PR.

[lahma]

Hey @tonyqus I believe this is ready for review. I've updated the benchmarks table to reflect latest version. I'm quite confident that this PR shouldn't break much and public API only has additions of ROS version side-by-side with the string-taking version. I was a good thing to have the tests as part of PR as they discovered one case I hadn't covered!

To improve read/load speed I think it would require changing from XmlDocument to XmlReader which is a lower-level API. I think there's some easy wins in Write part though, there's some unnecessary string allocations there.

[tonyqus]

I did think of changing XmlDocument to XmlReader. But it's kind of tradeoff between convenience and performance. XmlDocument is much easier to use than XmlReader for most cases.

[tonyqus]

Can you help put CellReferenceParser.cs into SS/Util folder instead of main folder?

[lahma]

Rebased and tweaks in last commit, Linux build still fails due to the fix PR still pending.

[tonyqus]

LGTM