Handle Memcached "object too large" consistently outside @cache

[allen0099]

Split out of #228 (fixed for @cache in #259).

Problem

Memcached rejects values larger than its item size limit (1 MB by default, memcached -I) with MemcacheServerError: object too large for cache. #259 makes @cache log this and serve the response unstored. Every other writer still gets the raw pymemcache error:

• CacheManager.set / add / get_or_set
• StateManager.create_state (large metadata)
• SessionManager (large session data)
• direct backend.set / set_if_absent callers

Redis has a much larger limit (512 MB per value), so the same code can work on Redis and fail on Memcached.

Options

1. Keep raising, with a clearer error. Translate the error into a CacheXError subclass that names the key's size and the server limit. This is the least surprising option for writers whose value must be stored: sessions and states must never be dropped silently.
2. Skip the store for pure caches. CacheManager.set could log and skip like @cache, since a cache miss later is harmless. get_or_set would still return the computed value. add would need a decision, because it reports whether it stored.
3. Check the size before sending. Add a max_item_size option on MemcachedBackend so the check happens without a round trip.

A reasonable split is 1 for sessions, states and the backend API, and 2 for CacheManager. Whatever is chosen, BACKENDS.md should document it per caller. The limit itself is already listed there as of #259.