Use global NonNullable type for nonnull types

[weswigham]

Also changes NonNullableAssertions to no longer be constraint locations, as it is no longer required to work with the appropriate members/type relationships with this change.

This PR uses the global NonNullable type alias if it can be found, and failing finding it (ie, when an old lib is used), it will manufacture an anonymous conditional type of the appropriate shape to use in its place (which will behave the same way, it just won't be pretty in quick info or declaration emit).

Fixes #21317

[sandersn]

1. How fast is this? Seems like it could be faster because of the caching, or slower because of the work in instantiating conditional types.
2. This seems like a decent solution to NonNullable being missing, while still allowing user overrides. But our other built-ins that come from lib.d.ts don't behave this way yet. The worst are Array and Promise, which don't error but have bad behaviour. As part of the general push to switch the compiler to rely more on conditional types from lib.d.ts, we should revisit our handling of types like this.

[weswigham]

1. I think probably faster (instantiation cache is a good thing, since it also causes relationships to be cached and stuff)... but do any of our perf tests have strictNullChecks on, even?

[sandersn]

No, although turning it on and obtaining a baseline from master would work, I think. That's what I did to test #22006.

[weswigham]

In my unscientific (ie, 10 iteration) tests, I'm not seeing any real difference just on the codebase using strictNullChecks. I think this is a very pay-for-play change - you only do much more expensive things when you're in a situation where more complex behavior is needed to behave correctly (eg, generics), which just don't happen often in your average (or at least our) codebase (but are still frustrating when they don't work) - usually just a handful of places. And the potential increase in work in the non-generic case (since the assignability checks done in getConditionalType are more complex than the checking filterType does) is clearly at least offset by the presence of an instantiation cache.

Metric	master	nonnull-not-constraint-position	Delta	Best	Worst
Compiler - Unions - node (v9.5.0, x64)
Memory used	208,801k (± 0.02%)	208,790k (± 0.05%)	-11k (- 0.01%)	208,554k	208,921k
Parse Time	0.68s (± 2.28%)	0.68s (± 1.41%)	-0.01s (- 1.02%)	0.66s	0.70s
Bind Time	0.69s (± 7.00%)	0.64s (± 3.06%)	-0.05s (- 7.22%)	0.61s	0.71s
Check Time	17.58s (± 1.37%)	17.47s (± 2.06%)	-0.10s (- 0.59%)	16.99s	18.76s
Emit Time	2.15s (± 0.99%)	2.13s (± 2.49%)	-0.01s (- 0.65%)	2.06s	2.30s
Total Time	21.10s (± 1.27%)	20.92s (± 1.97%)	-0.18s (- 0.85%)	20.36s	22.40s
Compiler - Unions - node (v9.4.0, x64)
Memory used	208,792k (± 0.03%)	208,835k (± 0.03%)	+43k (+ 0.02%)	208,661k	208,983k
Parse Time	0.67s (± 0.87%)	0.66s (± 0.89%)	-0.00s (- 0.45%)	0.65s	0.67s
Bind Time	0.64s (± 1.05%)	0.63s (± 0.64%)	-0.01s (- 1.26%)	0.62s	0.64s
Check Time	17.31s (± 0.50%)	17.28s (± 0.36%)	-0.03s (- 0.17%)	17.10s	17.39s
Emit Time	2.14s (± 1.14%)	2.15s (± 2.57%)	+0.01s (+ 0.47%)	2.08s	2.30s
Total Time	20.75s (± 0.48%)	20.72s (± 0.31%)	-0.03s (- 0.14%)	20.53s	20.85s