The short answer is: all things being equal, rasterize then transform.
The long answer depends on several factors:
- The size of the feature
- The allowable error
- Whether the CRSes agree in conformality and to what degree
- Which ellipsoid and datum are used
- Processing time constraints
The problem is that straight lines in one CRS won't necessarily be straight in another - unless they're both conformal - so transforming a line segment will position its end points accurately, but no new points will be added in between, so the line as a whole won't be accurate.
In your case, if you transform then rasterize, and the features are smallish in size, generally covering less that a square kilometre or so, there will be no appreciable error at any reasonable resolution of raster. But as you head towards the poles, or as your features get bigger, you'll notice greater error. So there will come a point at which the accuracy of your source data is better than the accuracy of your projected data - your allowable error.
Computationally it's generally quicker to transform vector data before rasterizing it, so you would have to weigh that up if you have any time constraints.
If you have time for some experimentation, I would generate two rasters using a subset of your data - one using the transform/rasterize process, and one the other way round. Then combine the two bitmaps with an exclusive-or operation, which will highlight areas where there is greater than one pixel error, and you can make a judgement on whether any error is allowable. You can also compare time taken for processing to further inform your decision.