The code that calls observers attempted to protect itself from the
list being mutated while it was iterating, but apparently it didn't
work well enough.
I changed the code so it now iterates the list collecting the
observer pointers, and then calls them afterwards.
While I was debugging this I noticed SequenceTracker doesn't have a
logging identifier, so I added it.

- Fixed a bug where the existing `split` method would report an
  extra empty string, if the input string ends with the delimiter.
- Added a variant of `split` that returns a vector instead of using
  a callback.

If a response had neither a Content-Length, Transfer-Encoding nor
Connection header, the error would be "Unsupported 'Connection'
response header", which is misleading.

Use Encoder::writeFormatted to simplify creating JSON.
Also used the not-at-all-new `enc["key"] = value` form.

Server was locking its mutex while calling a handler. This made it
impossible to handle more than one HTTP request at once. Especially
bad when handling `/db/_replicate`, whose handler blocks until the
replication completes!

Body::bodyAsJSON() would silently return nullptr if the response is
supposed to be JSON but fails to parse. This made some test failures
hard to track down today! It now logs a warning.

- Added "Server:" response header
- Added "ETag:" response header to document response
- Honor "If-None-Match" request header getting document
- Added "total_rows" and "update_seq" to _all_docs response
- Suppress response body on HEAD request
- _changes feed improvements:
  - Flush output before waiting
  - No heartbeat if ?heartbeat=0
  - Longest possible timeout if ?timeout=0
  - Stop on socket error
  - Make logging verbose not info

Instead of writing them immediately, save them in a Headers object.
That gives client code more flexibility about when it can set
headers -- it doesn't have to wait until after the status line is
sent.

It doesn't make a difference functionally, but it works around curl
not displaying the last line of output.

Discovered this scenario while testing replications in liteserv:
1. Create an empty database
2. Close it
3. Reopen it without giving the kC4DB_Create flag
4. Access the BlobStore, i.e. call c4Database::getBlobStore()

Result: getBlobStore fails with a NotFound error.
The reason is that step 1 didn't instantiate the BlobStore or create
the Attachments directory; and then in step 4 the db's flags don't
include kC4DB_Create so the BlobStore instance won't create the
directory.

Fixed by passing the kC4DB_Create flag to the C4BlobStore constructor
whenever the database is writeable.

- Moved all the request parsing out of the Task class.
- One-shot replication does not block response until finished.
- Returns 202 Accepted status, not 200.
- To cancel a replication you now set the `cancel` property to the
  `session_id` of the task.
- Preliminary support for channel/docID filters, but they're not
  per-collection yet.

Client specifies the version in a header, e.g:
  API-Version: 1
  API-Version: 3.5
If missing, it defaults to 1.0.
Handlers support a specific major version, and an incompatible
version in a request triggers a 400 response.

CBL will probably never use this, but liteserv needs it.

They're slow, and we only need very simple pattern matching.

Rather, get it when accepting the socket connection and reuse it.
Yes, this was actually a significant time-sink, roughly 6% of the
time to handle a simple request.

There were issues making it work with the DatabasePool.

Added `serverName`, `serverVersion` to C4ListenerConfig.
By default these should be nullslice.