Hi Jun,
that command line, which invokes the default tree inference, gives you an unrooted tree. You only infer explicitely rooted trees, when using an asymmetric model (e.g. binary Dollo, only change possible is from 0 -> 1). Any GTR+Gamma model implementation always leads to an unrooted tree. The "outgroup" option only roots the inferred unrooted tree by assuming KT005800 is the outgroup (the graphic manipulation, Alexi mentioned), i.e. places the root on the KT005800 terminal branch.
Which is what we call a trivial split: KT005800 | all other tips. It's called a "trivial split" because any terminal internode, the connection between a tip and the rest of the tree, must be in all bootstrap pseudoreplicates, it's BS support is always 100. Only the internodes seperating two or more tips from all others, i.e. the splits that divide the taxon set into two subset, hence, also called taxon bipartitions, are optimised and gauged during tree inference. Note that to establish BS support, we simply count the biparitions across the pseudoreplicate tree samples.
FigTree, being designed for dated trees, which are ultrametrised, hence, have an implicit root without the need to define an outgroup, interpretes the internode support as the support for the next higher node in the (here) outgroup-rooted tree. Here's a quick to the point for your tree.

But BS support are always branch support (internode) values, never node supports (common error in phylogenetic literature): a node in a tree is the joining point of tree internodes, i.e. depedending where I place the root, I propagate the support of one of the three internodes on the connecting node when viewing a standard NEWICK tree in (most) tree viewer(s) and assuming one or several tips represent the sister lineage, the outgroup. Placing BS support on nodes is an interpretation, not a result. PS this is also the reason why FigTree allows you to view the branch support on branches as the primary result (click-down menu: branch labels). BTW, the same holds for posterior probabilities, they are also inferred branch support, not node support!
The critical bipartition regarding the deep relationships in your data set is the violet split with the ambiguous BS support of 57: outtaxon tip + blue species | remaining ingroup. Noting the extreme phylogenetic distance (i.e. via the tree's branches) between all of the ingroup and the (single-tip) outgroup compared to the phylogenetic distance between the three "species" groups (could be populations, but genetically they would be species being high-coherent; here near-invariable); the low support for this critical bipartition (or internode, or branch) could be due to ingroup-outgroup long-branch attraction: ML has a 50% chance to escape LBA in the so-called Felsenstein Zone, and it well may be that it's 57 to 43 for the LBA-triggered purple split vs. a non-LBA biased alternative. Or it may be that some BS pseudoreplicate dataset suffer less from ingroup-outgroup LBA than others. Or that it is the genuine root, and the greens are sisters to the rest, but some BS pseudoreplicates (the other 43%) suffer from LBA.
You can visualise the alternative in the BS pseudoreplicate sample by reading in RAxML's bootstrap sample into e.g. SplitsTree or using the Networx object implemented in R/phangorn.
See this open access paper:
Cheers, Guido.