1.Would it be a potential bias if I have 2 replicates for one condition and 4 replicates for the other? Or would it better to use then only 2 replicates per condition?

Dataset 1:
cond A,  2 replicates
cond B, 4 replicates
 

2.Also, in my another dataset I have 2 replicates extracted with Long PolyA+ RNA (single end, 50 bp) and 2 replicates (single end, 36 bp) extracted with PolyA+RNA and 2 replicates with Long PolyA+RNA (paired reads, 76) for condition A, the same for condition B. Would it be possible in design matrix to combine the replicates in order to have 6 rep for condition A and 6 for condition B? Or the replicates from different extraction methods are better to be analyzed separately?

Dataset 2:
cond A, Long PolyA+ RNA, single end, 50 bp, 2 rep
cond A, PolyA+RNA,           single end, 36 bp, 2 rep
cond A, Long PolyA+RNA, paired end, 76bp,  2 rep

cond B, Long PolyA+ RNA, single end, 50 bp, 2 rep
cond B, PolyA+RNA,           single end, 36 bp, 2 rep
cond B, Long PolyA+RNA​, paired end, 76bp,  2 rep

For your first question, no. All popular methods for DE analyses can handle unbalanced designs.

For your second question - is it possible? Yes. Is it advisable? That's harder to answer. If the only aspect that differed between samples was the read length, then combining the samples might be okay, provided all the reads were long enough to be mapped accurately and counted into genes. However, differences in read length are often symptomatic of differences elsewhere, e.g., in sequencing batch, in library preparation, or in the processing center. You'd probably get a lot of protocol-specific biases that would inflate the variability if you treated them as direct replicates. If you're going to analyze them together, I would at least block on the extraction/preparation/sequencing protocol to mitigate these effects.