Add configurable alignment to FastSerializer

[josalem]

IOStreamStreamReader had a hard coded 8 byte alignment. EventPipe streams do not have any alignment requirements (save for some 4 byte alignments inside blocks). This means that blocks can have sizes that are not aligned. EventPipeEventSource will not dispatch events read from the stream until reads complete. In live cases where these blocks are unaligned, but no new data is coming down the pipe, the reader can attempt to enforce the alignment and end up blocking until new data arrives. This results in latency for event dispatch. For infrequent events, this can be a long period of time causing users to think they haven't received events.

For example:

EventBlock 1, EventBlock 2, and EventBlock 3 all arrive with aligned lengths. EventBlock 4 arrives with an unaligned length. Until more data arrives on the stream, none of the events in EventBlock 4 will be dispatched to the user.

This change adds a configuration switch for changing the alignment of the reader. The default is the old behavior (8 byte alignment) and the live reading for EventPipe was changed to 1 byte alignment.

Changing to an alignment of 1 uncovered a bug where calling Goto(StreamLabel) with a label greater than the alignment length would cause the subsequent read to return bytes from the wrong position in the stream. Up until now, this wasn't an issue for EventPipe because all Goto calls were for jumps < 4 due to the internal alignment, which is always less than 8.

This patch also fixes that bug.

TODO: tag all the issues this will close. I believe there are issues in dotnet/runtime, dotnet/diagnostics, dotnet/dotnet-monitor, and microsoft/perfview that are tracking this issue.

CC @tommcdon @jander-msft @wiktork

[noahfalk]

LGTM

[brianrob]

Overall, LTGM. A couple of small things.