Avoid allocations in more cases

[rynowak]

Updates to DFA Matcher to avoid allocations in more cases. This makes
the matcher more pay-for-play.

• Avoid allocating an RVD while matching if possible
• Avoid allocating the candidate set unless necessary

First, avoid creating the RVD unless there are parameters or
constraints. This means that the candidate set can contain null route
value dictionaries. This seems fine because selectors are already
low-level. The route values feature will allocate an RVD when accessed,
so code in MVC or middleware won't even see a null RVD.

Secondly, avoid creating the candidate set unless there are selectors.
This will save an allocation most of the time since we won't need to run
selectors is really common cases. The candidate set is needed because
selectors mix mutability and async, so it needs to be a reference type.
However the default case is that we don't need to run selectors. The
impact of this is that we make a bunch of methods have an instance
variant and a static variant that operates on the array.

---

I ran some perf tests on this, and it helps the case where you have multiple endpoints that match but no policies. We don't have any tests that directly benchmark this scenario, but it's somewhat common when you use parameters and literals in the same position, or use constraints.

Numbers from GithubMatcherBenchmark - this benchmark has some cases like this, and there's a marginal improvement across the aggregate.

Before

BenchmarkDotNet=v0.10.13, OS=Windows 10.0.17763
Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R), 1 CPU, 8 logical cores and 4 physical cores
.NET Core SDK=3.0.100-preview6-011546
  [Host]     : .NET Core 3.0.0-preview6-27623-04 (CoreCLR 4.6.27622.74, CoreFX 4.700.19.22212), 64bit RyuJIT
  Job-VZXIJQ : .NET Core 3.0.0-preview6-27623-04 (CoreCLR 4.6.27622.74, CoreFX 4.700.19.22212), 64bit RyuJIT

Runtime=Core  Server=True  Toolchain=.NET Core 3.0  
RunStrategy=Throughput  

Method	Mean	Error	StdDev	Op/s	Scaled	ScaledSD	Gen 0	Allocated
Baseline	147.4 ns	2.075 ns	1.839 ns	6,783,660.9	1.00	0.00	1.7090	9.49 KB
Dfa	354.7 ns	4.589 ns	3.583 ns	2,819,073.1	2.41	0.04	10.8643	57.07 KB

After

BenchmarkDotNet=v0.10.13, OS=Windows 10.0.17763
Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R), 1 CPU, 8 logical cores and 4 physical cores
.NET Core SDK=3.0.100-preview6-011546
  [Host]     : .NET Core 3.0.0-preview6-27623-04 (CoreCLR 4.6.27622.74, CoreFX 4.700.19.22212), 64bit RyuJIT
  Job-IRODAN : .NET Core 3.0.0-preview6-27623-04 (CoreCLR 4.6.27622.74, CoreFX 4.700.19.22212), 64bit RyuJIT

Runtime=Core  Server=True  Toolchain=.NET Core 3.0  
RunStrategy=Throughput  

Method	Mean	Error	StdDev	Op/s	Scaled	ScaledSD	Gen 0	Allocated
Baseline	144.1 ns	1.231 ns	1.152 ns	6,940,674.8	1.00	0.00	1.8311	9.49 KB
Dfa	338.6 ns	2.982 ns	2.490 ns	2,953,168.6	2.35	0.02	9.5215	52.12 KB

[davidfowl]

I approve this message

[davidfowl]

What allocations are left? Just the selector context?

[rynowak]

What allocations are left? Just the selector context?

Selector context - AND the CandiateState[] that we allocate.

[rynowak]

@aspnet-hello triage https://dev.azure.com/dnceng/public/_build/results?buildId=171062