The topicmodels package provides an interface to the GSL C and C++ code for topic models by Blei et al. and Phan et al. For the earlier it uses Variational EM, for the latter Gibbs Sampling. See http://www.jstatsoft.org/v40/i13/paper . The package works well with the utilities from the tm package.
The lda package uses a collapsed Gibbs Sampler for a number of models similar to those from the GSL library. However, it has been implemented by the package authors itself, not by Blei et al. This implementation therefore differs in general from the estimation technique proposed in the original papers introducing these model variants, where the VEM algorithm is usually applied. On the other hand, the package offers more functionality then the other package.
The package provides text mining functionality too.
Regarding extensibility, the topicmodel code by its very nature can be extended to interface other topic model code written in C and C++. The lda package seems to be more relying on the specific implementation provided by the authors, but there Gibbs sampler might allow specifying your own topic model. For extensibility issues nota bene, the former is licensed under GPL-2 and the latter LGPL, so it might depend on what you need to extend it for (GPL-2 is stricter regarding the open source aspect, i.e. you can't use it in proprietary software).
I can't help you here, I only used topicmodels so far.
Personally I use
topicmodels, as it is well documented (see the JSS paper above) and I trust the authors (Grün also implemeted flexmix and Hornik is R core member).