Eliminate HTTP2 HPack enumerator allocations

[JamesNK]

Fixes #19275

Creates a shared struct "enumerator" that wraps the two possible struct enumerator types (avoids boxing to IEnumerator) and handles enumerating multiple strings in StringValue (avoids allocation from yield).

Long term we should improve the way ASP.NET Core does HPack and try to unify more with the client, but in the meantime this change eliminates 4 allocations per HTTP/2 request.

There are some minor changes to shared HTTP2 code that will need to be mirrored.

[JamesNK]

Before:

After:

[JamesNK]

I've changed this PR to use an #if, and our code continues to use HPackEncoder directly. Reasons:

• It minimizes the amount of overall change.
• HttpClient might not use HPackEncoder.BeginEncode but non-Kestrel code in our repo does, e.g. HttpSys tests, httpcat.

[JamesNK]

Yield (current)

               Method |     Mean |    Error |   StdDev |        Op/s |  Gen 0 | Allocated |
--------------------- |---------:|---------:|---------:|------------:|-------:|----------:|
 WriteResponseHeaders | 604.2 ns | 3.755 ns | 3.512 ns | 1,655,046.7 | 0.0029 |     232 B |

Struct enumerator

               Method |     Mean |    Error |   StdDev |        Op/s | Allocated |
--------------------- |---------:|---------:|---------:|------------:|----------:|
 WriteResponseHeaders | 664.2 ns | 6.750 ns | 4.881 ns | 1,505,680.8 |      32 B |

Class enumerator as IEnumerator

               Method |     Mean |    Error |   StdDev |        Op/s | Allocated |
--------------------- |---------:|---------:|---------:|------------:|----------:|
 WriteResponseHeaders | 568.8 ns | 3.868 ns | 3.618 ns | 1,758,145.7 |      32 B |

Class enumerator

               Method |     Mean |    Error |   StdDev |        Op/s | Allocated |
--------------------- |---------:|---------:|---------:|------------:|----------:|
 WriteResponseHeaders | 526.4 ns | 5.163 ns | 4.829 ns | 1,899,719.1 |      32 B |

A cached enumerator on the frame writer and HPackEncoder taking that type directly (avoid interface calls on IEnumerator) is fastest.

[davidfowl]

LGTM