The reason for the first question about replaces and re-invites is that with most providers and situation, this should be on. It allows the calls to be transferred and other features that would normally put the call on-hold while the transfer or call handling is negotiated and implemented. The media server is indeed the MOH source as I indicated.
Assume a rule of thumb of 100kps for voice using g711 in each direction. Even with 128 calls your demand for bandwidth is pretty minimal for voice. Of course, this is dependent on your other "data" traffic. Let's not forget that internal calls may not require transcoding at the system level as the phones will accommodate provided that you leave PBX delivers Audio off at the extension level.
While you may have g711a set up, what if an outside call comes in and that call is using G711u? G711a seems to be more widely used in EU while g711u is more prevalent in the US. Your provider most likely supports both as well as G729 and perhaps others. Each leg of a call may require some negotiation and transcoding and I think it possible that your provider will still attempt to pass the originating codec type to you provided that your set-up will support. I assume you have the 3CX system set with G711a at the top with others following and that the same was done for each phone connected. If so, then your system will likely accept the offered codec.
You can set PBX delivers audio set at the provider level and/or at the phone level. I always encourage at the provider level. I do not at the phone level as the phones can handle it for themselves (extension to extension for example) as I indicated above.
The issue you are concerned about is latency. Each aspect (analog to digital conversion at the phone, data transmission, data packet handling at the router and switches, transcoding, VoIP provider handling, PSTN network, etc.) of the call handling process to its destination and back adds to the overall latency. Usually the worst latency comes when forwarding using a cell/mobile number as that leg has to encounter more processing due to the network challenges of further call handling and protocol conversions (GSM, TDM or CDMA).
Transcoding is generally not a big factor, but as you suggested, the more calls - the possibility exists that the processor becomes more bogged down and therefor adds more to the latency issue. The real issue is whether you have a processor up to the task. Search the blog as 3CX has published a recommended system minimum configuration based on calls.
http://www.3cx.com/downloads/3CX_Phone_System_Operating_System_Specifications.pdf
In the scheme of things, transcoding is but one piece of the overall latency, but my take is that when call forwarding and then possibly to a cell phone as well, the latency that transcoding interjects is minimal. The vast majority is network related - that is to get the call from point A (phone to 3CX) to B (3CX to provider) and then from B to C (provider to PSTN) and if cell from C to D and even possibly D to E (depending upon end cell provider) and back again.
While I am surprised that the ping time is lower to the provider than to your own sites, I attribute the delay perhaps to the efficiency of any VPN?
I encourage you (again) to use wireshark if you really want to examine the flow. You can then experiment and have an objective measure by which you can then decide what settings and all suit your environment best. I also encourage you to search the blog as there have been a number of studies that deal with some of what you are asking.
