[npazosmendez]

Most Graphite Web's graph functions (color(), threshold() etc) "work" with non-image formats (e.g. format=json) by being mostly no-op (depends on the function). For example, with color(randomWalk('a'), 'red'):

$ docker run -d --name gw -p 8080:80 graphiteapp/graphite-statsd >/dev/null && until curl -sf -o /dev/null "http://localhost:8080/render?target=x&format=json"; do sleep 2; done && curl -sG http://localhost:8080/render --data-urlencode "target=color(randomWalk('a'), 'red')" -d format=json -d from=-5min; echo; docker rm -f gw >/dev/null

[{"target": "a", "tags": {"name": "a"}, "datapoints": [[0, 1785953431], [0.24124975089820433, 1785953491], [-0.03474284243293879, 1785953551], [-0.47080131891509025, 1785953611], [-0.3820169792063769, 1785953671]]}]

This behavior is useful for two reasons: 1) some functions make sense even without image rendering (e.g. threshold), 2) queries don't completely break when switching formats

Carbonapi follows that behavior, but only when built with cairo. When built without cairo, they return no data (or an error in some cases).

This PR changes that, and allows such functions to run correctly in no-cairo builds if possible. Only the image generation bit (which depends on the format= requested) remains under the cairo build tag. If built without cairo, an error is returned, but only when an image is requested explicitly.

Behavior changes:

• Non-cairo builds: color, stacked, areaBetween, alpha, dashed, drawAsInfinite, secondYAxis, lineWidth, threshold, verticalLine and holtWintersConfidenceArea now evaluate normally and appear in /functions.
• Non-cairo builds: format=png/svg returns 501 instead of a 200 with an empty body. Notably, since png is the default format, requests without format also get 501 now.

Other changes:

• CI now runs the existing test_nocairo
• Perf: because the GraphOptions struct does not depend on the build tag, MetricData type size grows in no-cairo builds from 192 bytes to 272 bytes. I think this is acceptable and will probably be negligible compared to the memory that comes from tags and sample values. Still I wanted to state it explicitly