Add cryptographic operation counts to prevent process crashes

[vcsjones]

OpenSSL, CNG, and CommonCrypto are not thread safe, and some improper uses result in hard process crashes. This is easiest to hit in paths that use initialize and reset primitives because the internal init routine frees and re-creates structures that might be in use by other operations. This has been observed in all three of our desktop platforms.

This pull request adds counts to in-flight operations for Hash and HMAC primitives so that we can throw managed exceptions in this scenario instead of process crashes.

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-security, @bartonjs, @vcsjones
See info in area-owners.md if you want to be subscribed.

[vcsjones]

Assuming this made it past the _running guard, this

1. Disposed the current handle in DestroyHash
2. Nulled out _hHash
3. Created a new hash
4. Assigned _hHash to the new one.

If another thread attempted to use that hash during 2 or 3, then some asserts would trip because we never expected _hHash to be null.

Now, we

1. Create a new hash.
2. Exchange old and new
3. Dispose of old.

This ensures we don't have a period of time where _hHash can be null and trip asserts during test runs.

[vcsjones]

Benchmarks basically showed no major change when the interlocked value is not under contention. I didn't both to benchmark when it is under contention since that is an error path.