robomimic-v1: the state-only variants (square, lift, can, transport) build

[tactino]

robomimic-v1 ships eight metadata files, but only the four -img ones could be built. square, lift, can and transport carry no camera_names, and RobomimicEnv read that key unconditionally, so each died with KeyError: 'camera_names'. The docs audit flagged it, and running it on guangzhao confirmed it. Every experiment used -img names, so nothing here had hit it.

The fix reads camera_names with an empty default. It also turns image observations on only when there are cameras. Turning them on with none would have robosuite add its default agentview image, and prepare_obs would have put it among the states. The agentview frame is rendered separately, as before, so every variant still has it.

Checked on guangzhao, in the robomimic environment E27, E34 and E40 ran in (robomimic 0.4.0, robosuite 1.4.1, MuJoCo 2.3.7):

name	before	after
square	KeyError: 'camera_names'	images agentview; 11 state keys
lift	KeyError	same
can	KeyError	same
transport	KeyError	images agentview; both robots' state keys
square-img	images agentview, robot0_eye_in_hand; the same 11 state keys	unchanged

CI does not install robomimic, so no test here can build the env. The table is the check.