Fix a goofy bug where the app has been using different key values for
storing data vs fetching data.

* Set higher than default resource limits
* Drop local kubectl usage
* Update shell command for upstream deprecation of `--serviceaccount`
* Fix debug command to actually work by changing working directory

Layers of caching can make things weird. We seem to be seeing crashes in
the app caused by nodes that are in the node list, but missing from the
metrics data.

Add a new "Monitoring" section to the page which will always contain
a link to a nice Grafana dashboard that Arturo has created. Also check
the namespace name and add links to the SAL tool and Striker for the
Tool account if that name starts with "tool-".

The Kubernetes client returns None rather than an empty list if there
are no active jobs.

Add a 'max_parts' parameter to the duration filter which limits the
number of segments (years, weeks, days, hours, minutes, seconds, millis)
used in the output. The default limit is 3, but callers can tune this up
or down as desired.

duration() allows the caller to provide an end datetime if period to
display is not relative to the current wall time. It also expands the
prior output of summarize() to include years and weeks.

Count pods in the 'Pending' and 'Running' phases as "active".

We have a small number of tools where no Pods are running, but the tool
is still "active" in that it handles URL routing. Show the namespaces of
these tools in the "active" namespace list too.

I read the source for HEAD, not the release that is in our venv.

The API return order is in sorted order on the node's name, which is
nice, but for 'reasons' we used non-padded numbers in cluster node
instance names. A "natural" sort uses numeric sequences as single keys
so that unpadded lists keep things grouped by orders of magnitude
instead of putting the digits in strictly alphanumeric order order.

This only shows up in the data if the ingress has been patched since
deploy.