Its always better to have a test done to ensure the results are accurate and reliable.

Yes its part of cfd validation as well as assure the accurate results

Grid independence study is essential to a CFD simulation. As mentioned by Mr Mehlhoop, good results can be incidental only. It is possible that by resolving the grid further you lose correlation with experimental results because THE FLOW IS BETTER RESOLVED. This would mean that the results of your earlier simulation were erroneous. Grid independence would provide the necessary confidence in the simulation to start studying correlation with experimental results for validation.

As the others researcher told you earlier, it is really important to find a grid dependancy to be able to fully justify your results. But I would like to add that it also depends on type of of experiemental data you're using: having an error of 0.5% on a global/mean coefficient is easier to have that 0.5% on your velocity field for exemple.

Anyway, if you want a good way to assess your mesh precision I already used this:

http://fluidsengineering.asmedigitalcollection.asme.org/article.aspx?articleid=1434171

Even if it doesn't work well in all cases, I recommend it since it is a really good way to compute your mesh error and know exactly which results you can use.

Comparison with experimental data and grid-independancy are two different concepts! Mathematically, you have to be ensured that your mathematical system (i.e. PDEs) has a "unique solution" not more! The extent of being close to that solution is something else!

grid indepence test is a necessary protocol which you have to follow while simulating a CFD problem. Getting correct results is not the sole purpose, decreasing computational cost by doing grid independence test is one of the most necessary part.

Yes of course. If you don't do a grid independence study you can have CFD results that match the experimental ones just due to luck. If you make the independence study you can asses that your numerical model is suitable for such kind of simulations/predictions and can be intensively used to study the phenomena without further experimental tests.