Truncated distributions define `.var` instead of `.variance`, so `.variance` raises NotImplementedError

[tillahoffmann]

Problem

Three public, exported truncated distributions define a .var property but not the standard .variance:

• LeftTruncatedDistribution — truncated.py:123
• RightTruncatedDistribution — truncated.py:211
• TwoSidedTruncatedDistribution — truncated.py:357

The canonical statistic across every other distribution (and the Distribution base class) is .variance. Because these classes only define .var, calling the standard .variance falls through to the base-class property, which raise NotImplementedError:

import numpyro.distributions as dist
d = dist.LeftTruncatedDistribution(dist.Normal(0, 1), low=0.0)
d.var       # -> works, returns the variance
d.variance  # -> NotImplementedError  (standard API is silently broken)

The .var bodies compute the truncated-Normal variance correctly — they're just exposed under a non-standard name, so the documented .variance API is broken for these three distributions.

Proposed fix

Rename var → variance on all three classes. This is safe and self-contained:

• No caller in the codebase reads .var on these distributions (the only .var references — in diagnostics.py and test_distributions.py — are numpy/scipy .var() calls, unrelated).
• Restores the standard .variance API.

Optionally add a test asserting .variance returns the value for these truncated distributions (the existing suite only exercised the wrongly-named .var indirectly, if at all).