rework numeric tokenizer hot path

[EliotJones]

the existing numeric tokenizer involved allocations and string parsing. since the number formats in pdf files are fairly predictable we can improve this substantially

this ends up being somewhere between 2-3 times faster in my benchmarks on a subset of numeric data from PDF files:

| Method | Mean      | Error     | StdDev    |
|------- |----------:|----------:|----------:|
| PigOld | 13.684 us | 0.2333 us | 0.2182 us |
| PigNew |  5.963 us | 0.0990 us | 0.0877 us |

| Method | Mean      | Error     | StdDev    |
|------- |----------:|----------:|----------:|
| PigOld | 14.806 us | 0.2098 us | 0.1962 us |
| PigNew |  6.230 us | 0.1205 us | 0.1127 us |

Based on tracing for opening ~350 documents the tokenize method here was called approximately 16.2 million times so is definitely a hot path. Real world performance impact may be different.

[EliotJones]

@BobLd I'm planning a 0.1.11 full release as follows:

• Merge make link copying more tolerant when adding page #1103 to improve link building
• Release full version after 1103 with various parsing improvements and bugfixes
• Merge this and move file parsing to single-pass static methods #1102 since these are much larger, more risky changes and will have time to bed-in as I work through the common crawl corpus and people trial pre-release versions
• Work on the builder redesign details in make link copying more tolerant when adding page #1103 for 0.1.12

[BobLd]

@EliotJones thanks a lot for this change, looks great!

Regarding 0.1.11, we are on the same page - agreed with your plan.