Bound concurrent image processing and upstream fetches (closes #64)

Nothing bounded total in-flight work, so a burst of cache misses across
hosts could exhaust memory. Two settings now do: max_concurrent_processing
(default the CPUs Go uses) and upstream_connections (default 64, beside
the per-host limit). A request that finds either full waits up to 10
seconds, then gets 503; a slot is released on every path. No request
holds source bytes while it waits: a cached source is read only after
the processing slot is taken, and a fetched one only while it holds its
upstream connection. libvips runs one worker thread per image with its
operation cache off. Both waits count toward downstream_timeout, as the
README says.

Model: opus-5-5

config.example.yml

@@ -71,6 +71,19 @@ allow_http: false
 # Maximum concurrent connections per upstream host (default: 20)
 upstream_connections_per_host: 20
 
+# Maximum concurrent connections to all upstream hosts together, on top of
+# the per-host limit (default: 64). A fetch holds its connection until its
+# image has been processed. A fetch that finds none free waits up to 10
+# seconds for one, and if none frees up the request is answered 503, unless
+# downstream_timeout has ended first.
+upstream_connections: 64
+
+# Maximum number of images decoded and encoded at once (default: the
+# number of CPUs pixa can use, which follows a container's CPU limit). A
+# request that finds none free waits up to 10 seconds for one, and if none
+# frees up it is answered 503, unless downstream_timeout has ended first.
+# max_concurrent_processing: 4
+
 # Time allowed for one fetch from an upstream host (default: 30s)
 upstream_fetch_timeout: 30s
 
@@ -78,8 +91,10 @@ upstream_fetch_timeout: 30s
 # (1 GiB) (default: 52428800, 50 MiB)
 upstream_max_response_size: 52428800
 
-# Time allowed for answering one client request, the upstream fetch
-# included, so keep it longer than upstream_fetch_timeout (default: 60s)
+# Time allowed for answering one client request (default: 60s). The
+# upstream fetch counts toward it, and so do the waits for an upstream
+# connection and for a processing slot (up to 10 seconds each), so keep it
+# longer than upstream_fetch_timeout plus 20 seconds.
 downstream_timeout: 60s
 
 # The origin a browser lets read pixa's responses, sent as the CORS